An Improved Line-Point Low-Degree Test

Prahladh Harsha∗ Mrinal Kumar∗ Ramprasad Saptharishi∗ Madhu Sudan†

arXiv:2311.12752v1 [cs.CC] 21 Nov 2023

Abstract

We prove that the most natural low-degree test for polynomials over finite fields is “robust”
in the high-error regime for linear-sized fields. Specifically we consider the “local” agreement of
a function f : Fmq → Fq from the space of degree-d polynomials, i.e., the expected agreement of
the function from univariate degree-d polynomials over a randomly chosen line in Fm q , and prove
τ
that if this local agreement is ε ≥ Ω(( /q ) )) for some fixed τ > 0, then there is a global degree-d
d

polynomial Q : Fm q → Fq with agreement nearly ε with f . This settles a long-standing open

question in the area of low-degree testing, yielding an O(d)-query robust test in the “high-error”
regime (i.e., when ε < 1/2). The previous results in this space either required ε > 1/2 (Polishchuk
& Spielman, STOC 1994), or q = Ω(d4 ) (Arora & Sudan, Combinatorica 2003), or needed to
measure local distance on 2-dimensional “planes” rather than one-dimensional lines leading to
Ω(d2 )-query complexity (Raz & Safra, STOC 1997).
Our analysis follows the spirit of most previous analyses in first analyzing the low-variable
case (m = O(1)) and then “bootstrapping” to general multivariate settings. Our main technical
novelty is a new analysis in the bivariate setting that exploits a previously known connection
between multivariate factorization and finding (or testing) low-degree polynomials, in a non
“black-box” manner. This connection was used roughly in a black-box manner in the work of
Arora & Sudan — and we show that opening up this black box and making some delicate
choices in the analysis leads to our essentially optimal analysis. A second contribution is a
bootstrapping analysis which manages to lift analyses for m = 2 directly to analyses for general
m, where previous works needed to work with m = 3 or m = 4 — arguably this bootstrapping
is significantly simpler than those in prior works.

∗
Tata Institute of Fundamental Research, Mumbai, India. [email protected], [email protected], ram-
[email protected].
†
School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA. Supported in part by a
Simons Investigator Award and NSF Award CCF 2152413. [email protected].

1
Contents
1 Introduction 3
1.1 Technical Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Our results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Proof overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Bivariate low-degree tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Bootstrapping to higher dimensions . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Organisation of the paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Preliminaries 13
2.1 Structure of minimal interpolating polynomials . . . . . . . . . . . . . . . . . . . . . 15
2.2 Power series roots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3 Low-degree roots on restrictions to global low-degree roots 16

4 The bivariate low-degree test 18

4.1 Some consequences of Theorem 4.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5 Lifting to m-variate low-degree tests 24

5.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.2 The bootstrapping statements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3 Low-error regime: Proof of Theorem 1.2 . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3.1 Self-correction passes LDT with better probability (Proof of Claim 5.9) . . . 27
5.4 High-error regime: Proof of Theorem 5.7 . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.4.1 Corrected function passes LDT with high probability (proof of Lemma 5.11) . 31

A An Arora-Sudan style interpolation 40

A.1 Finding a structured subset of points . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
A.2 Properties of the desired interpolating polynomial . . . . . . . . . . . . . . . . . . . . 43
A.3 Proof of the interpolation lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
A.4 Proofs of Claim A.4, Claim A.5 and Claim A.6 . . . . . . . . . . . . . . . . . . . . . 46

B Variations on Low-degree Testing 48

2
1 Introduction
In this paper we consider the classical “line-point” “low-degree test” for multivariate polynomials
over finite fields and give a “near-optimal” analysis in the “high-error” regime. We expand on these
terms below.
The basic question in “low-degree testing” is to estimate the distance of a function f : Fm
q → Fq
given as an oracle from the space m-variate polynomials over Fq of degree at most d. Broadly,
for a family of function F ⊆ {f : D → R}, a t-query α(·)-robust test for F is a distribu-
tion supported on sets S ⊆ D with |S| ≤ t such that for every function f : D → R we have
δ(f, F ) ≤ α(ES [δ(f |S , FS )]). (Here we use δ(f, g) to denote the normalized Hamming distance
between functions, δ(f, F ) := minP ∈F {δ(f, P )}, f |S to denote the restriction of f to the domain
S and F |S to denote {P |S |P ∈ F }.) The quantity ES [δ(f |S , FS )] is thus a local measure of the
distance of f from F and α(·)-robustness relates this local distance to the global distance of f
from F . Low-degree testing is the specialization of the study of robust testing in the case where
F = Fdq [X1 , . . . , Xm ], namely the space of m-variate polynomials of degree at most d over Fq ,
viewed as functions from Fm
q to Fq . The error (tolerance) of a test is roughly this highest value of δ
such that α(δ) is (noticeably) bounded away from 1. (Specifically in the case of low-degree testing
p
we would want α(δ) < 1 − d/q.) For most natural tests, it is straightforward to actually show
α(δ) ≈ δ as long as it is noticeably bounded away from 1. So the critical parameter desribing α
is just the error-tolerance and this is what we will use to describe the history (and importance) of
low-degree testing.
The low-degree testing problem is a classical problem with enormous impact in the early as well
as state-of-the-art constructions of probabilistically checkable proofs (PCPs). It was introduced by
Rubinfeld and Sudan [RS96] who, in our language, gave a q-query test that had an error-tolerance of
O(1/d) provided q = Ω(d2 ). The test they introduced is now called the “lines-point” test and is given
by the uniform distribution on lines in Fm
q . Arora, Lund, Motwani, Sudan and Szegedy [ALMSS98],
building on the work of Arora and Safra [AS98], improved the error-tolerance of the lines-point test
to Ω(1) (provided q = Ω(d3 )) and this was a crucial ingredient in getting PCPs with O(1)-query
complexity. Subsequent improvements to PCP parameters were also closely related to improvements
to (analyses of) low-degree tests. For instance the first constructions of nearly-linear sized PCPs
by Polishchuk and Spielman [PS94] rely on getting error = Ω(1) for q = O(d) and m = O(1) in the
lines-point test. This was extended to general m by Friedl and Sudan [FS95] which till this work
remains the optimal analysis for the lines-point test when q = O(d).
Higher error low-degree tests and analyses, with error tending to 1, were obtained by Raz
and Safra [RS97] and Arora and Sudan [AS03]. The former introduced the “planes-point test”,
where the underlying distribution is uniform on “planes” (i.e., 2-dimensional affine subspaces of
Fm τ
q ) and showed that it had error 1 − ( /q ) for constant τ . Thus, this obtains essentially optimal
d

error, but at the cost of q 2 -queries. The latter ([AS03]) improved the analysis of the line-point

3
test, but only in the regime where q was super-quartic in d and showed roughly that it had error
1 − (d4/q )Ω(1) . So this reduced the query complexity to O(q), but only when q = Ω(d4 ). Thus,
the three results above are essentially incomparable and represent the the three state-of-the-art
low-degree tests today. (We remark there is also a vast body of related questions starting with the
work of Alon, Kaufman, Krivelevich, Litsyn and Ron [AKKLR05] that might be termed “moderate
degree testing” where one considers the setting q < d ≪ m. These results and their motivations are
quite distinct from those in this work and we do not cover those results here. A direction of study
more related to the setting of this paper (i.e., when d ≪ q) involves derandomizing the low-degree
test [BSVW03, MR08]. This direction turns out to be crucial in getting PCPs of small (near-linear)
size [BGHSV06, BGHSV05, MR10a, MR10b]. We do not pursue this direction in this paper, though
it can be a subject of further study.)
The main result of this paper is a single analysis of low-degree testing that qualitatively subsumes
all previous low-degree tests. We analyze the lines-point low-degree test and give an analysis showing
it has error tolerance 1 − (d/q )Ω(1) when q = O(d) (See Theorem 1.5).

1.1 Technical Contributions

Most analyses of low-degree tests follow the following paradigm: One first analyzes the low-degree
test in the setting of m = O(1), i.e., with a constant number of variables. And then a second step
of analysis “bootstraps” the result from O(1)-variables to general m variables. Our improvement
follows the same paradigm and contributes to both steps. We explain our contribution to the two
steps below by contrasting with the previous works.
Previous techniques in the O(1)-variable setting come in two distinct flavors: The Raz-Safra
analysis [RS97] is very coding theoretic. For instance when testing 3-variate functions by picking
planes uniformly, the analysis relies on the fact that two typical planes intersect in a line, and on
this line the nearest polynomial is a codeword of a code with 1 − o(1) distance (if d = o(q)). This
very high distance of the underlying code is critical to their analysis. The Raz-Safra analysis is thus
very clean, but it simply can not work with the lines-point test where typical lines intersect in at
most one point.
The O(1)-variable analyses for the line-point tests in [PS94] and in [AS03] are both very algebraic.
Both rely on ideas used to decode Reed-Solomon codes, but need some heavy machinery to work with
these. [PS94] in particular use properties of the derivatives of resultants to effect their analysis;
and [AS03] uses effective Hilbert irreducibility to effect their analysis. While such use of heavy
machinery is inevitable given the approach, the weakness in previous results comes from the black
box use of the tools that they use. For instance [PS94] effectively only uses the fact that when
m = 2, the low-degree test effectively gives two directions in which the function looks like a low-
degree polynomial in that direction, but is unable to use the fact that Ω(1)-fraction of directions
actually have this property. [AS03] also suffers from a black box use of Hilbert irreducibility. Our

4
key contribution (and we elaborate more on this in Section 1.3) is to open up this black box and
adapt it to our setting. (Our proofs as a result are self-contained, and arguably simpler than those
in [AS03].)
Turning to the “bootstrapping” — here there are roughly three previous works to compare
against. The Raz-Safra bootstrapping [RS97] turns out to be the weakest and shows that an error
tolerance of 1 − ε in the O(1)-variable setting can be converted to an error tolerance of 1 − O(mε) in
the m-variate setting. While this may be adequate in some settings, this is certainly not the right
answer. In the low-error setting, Friedl and Sudan [FS95], essentially building on Rubinfeld and
Sudan [RS96], show that an error upper bound of ε in the 2-variable setting can be converted to an
error bound of ε/C (for some universal constant C) in the m-variate setting; but their proof is very
highly tailored to the unique-decoding setting, i.e., the proofs use the fact that in such settings there
is at most one polynomial at ε distance from any given function. Arora and Sudan [AS03] extend
this analysis to the list-decoding setting but their proof is complex and finally is only able to reduce
to the 3-variable case. Our analysis shows that an error-tolerance of 1 − ε in the 2-variable setting
implies a 1 − εΩ(1) error tolerance in the general m. While our analysis is in the same spirit as the
previous works, is substantially cleaner and manages to cleanly reduce to the 2-variable case. We
remark that our analysis is inspired by the clean local-list-decoder for multivariate polynomial codes
due to Sudan, Trevisan and Vadhan [STV01] who in turn clean up a similar result from [AS03], an
idea that does not seem to have been explored much in the interim period.
We remark that both the [FS95] analysis and the [AS03] analysis can be interpreted as exploiting
some expansion properties of underlying high-dimensional expanders (HDXs). The [FS95] HDX
consists of three layers, the points in Fm m
q , the lines in Fq and some complicated 2-dimensional
surfaces in Fm
q . The [AS03] HDX is simply the Grasmannian (points, lines, planes and cubes), but
now it has four layers. The cleanliness of our analysis is highlighted by the fact that we also work
with the (affine) Grassmannian in Fm
q , but now again with three layers (points, lines and planes).
Finally we remark that we do not optimize the dependence between q and the ε in the error-
tolerance we obtain here. Recent works due to Bhangale, Dinur and Navon [BDN17] and Minzer
and Zheng [MZ23] explore this connection and obtain near-optimal dependence between ε and q,
albeit at the cost of an even larger query test – a “cubes-point” test. However, these results do
not optimize the dependence on d. These results are proved by using deeper expansion properties
of the 4-layered Grassmannian (points, lines, planes and cubes) in Fm
q . If one obtains an optimal
ε-vs-q tradeoff for the lines-point test, with q = O(d) as in our proof, this might yield hardness of
labelcover of the following type: NP-hard to distinguish between labelcover instances with alphabet
size O(q 2 ) with perfect completeness from those that are at most O((log q)/q)-satisfible. The current
best result, in this context, is due Siu on Chan [Cha16], albeit with imperfect completeness. We
believe these questions merit further study and view our work as a step towards understanding
them.

5
1.2 Our results
To state our results which work in the high-error regime of the lines-point test, it will be best to first
recall similar results in the low-error regime. We begin with the following lemma due to Polishchuk
and Spielman [PS94].

Lemma 1.1 (Polishchuk-Spielman Lemma [PS94, Lemma 8], see also [BCIKS20, Lemma 4.3]).
Let F be any field and let A(x, y, z) ∈ F[x, y, z] be a non-zero trivariate polynomial with z-degree
at most 1 and (1, 0, d)-degree and (0, 1, d)-degree at most D. Furthermore, suppose there exist two
sets U, V ⊆ F such that (1) for every u ∈ U , there exists a degree-d univariate polynomial Cu (y)
such that A(u, y, Cu (y)) ≡ 0 and similarly (2) for every v ∈ V , there exists a a degree-d univariate
polynomial Rv (x) such that A(x, v, Rv (x)) ≡ 0.
If the sets U, V are of size at least 2D each then there exists a polynomial Q(x, y) of individual
degree at most d in each variable such that

A(x, y, Q(x, y)) ≡ 0.

Friedl and Sudan [FS95] used the above lemma to prove the following low-error lines-point theorem.

Theorem 1.2 (Multivariate low-error LDT [FS95]). There is a constant C large enough such that
for every finite field Fq , m > 0, degree d satisfying q > C · d, the following holds:

Suppose the points table f : Fm

q → Fq and the degree-d lines oracle Pℓ satisfy

Pr[f (a) 6= Pℓ (a)] ≤ δ,

a,ℓ

for some 0 < δ < 0.01, then there exists a m-variate degree-d polynomial Q such that
δ(f, Q) ≤ 4δ.

We are interested in proving high-error versions of the above results. A natural approach will
be to extend the Polishchuk-Spielman Lemma 1.1 to higher degrees in z. However, the natural
generalization happens to be false. Consider the trivariate polynomial

A(x, y, z) := (xz − C(y)) · (yz − R(x))

where C ∈ F[y], R ∈ F[x] are two polynomials of degree at most d. Clearly for each u 6= 0,
A(u, y, C(y)/u) ≡ 0 and for each v 6= 0, A(x, v, R(x)/v) ≡ 0. Yet, there exists no Q(x, y) such that
A(x, y, Q(x, y)) ≡ 0. This counterexample exists as z-degree of A is at least the number of parallel
directions1 . Our main lemma shows that if the number of parallel directions is considerably more
1
There exist similar counterexamples for every r ≥ 2 with z-degree being r and a set of r-parallel directions. The
above example with r = 2 has 2 sets of parallel directions (lines parallel to the x-axis and those parallel to the y-axis).

6
than the z-degree, then such counterexamples do not exist. We find it more convenient to state our
lemma for sets of lines passing through a point rather than sets of parallel lines. 2

We mention a slightly informal statement of our main lemma.

Lemma 1.3 (Main technical lemma (Informal)). Let F be any finite field and A(x, y, z) ∈ F[x, y, z]
be a non-zero trivariate polynomial with z-degree at most dz and (1, 1, d)-weighted degree at most D
such that the characteristic of F is greater than dz .
If there is a set B ⊆ F2 that satisfies (1) |B| > 2dz D|F|, and (2) for every (α, β) ∈ B, there
is a set S(α,β) ⊆ F2 of size greater than (dz D|F|) such that for every (u, v) ∈ S(α,β) , there exists a
univariate polynomial P((α,β),(u,v)) (t) ∈ F[t] of degree at most d for which


A α + tu, β + tv, P((α,β),(u,v)) (t) ≡ 0 ,

then, there exists a polynomial P (x, y) ∈ F[x, y] of total degree at most d such that

A(x, y, P (x, y)) ≡ 0 .

Remark. Even though Lemma 1.3 is stated for trivariate polynomials here, the statement is true
for multivariate polynomials as it is. Moreover, even though the underlying field F is assumed to be
finite here and with large enough characteristic, a similar statement that suffices for our applications
to low-degree testing is true over all fields. We refer to Section 3 for details. ♦
Given this lemma, we can now prove the following high-error bivariate lines-point theorem.

Theorem 1.4 (Bivariate low-degree test). There exists a constant τ ∈ (0, 1] such that for every
finite field Fq , ε0 ∈ (0, 1) and degree d satisfying ε0 > (d/q)τ , the following holds.
Suppose the points table f : F2q → Fq and degree-d lines oracle Pℓ satisfy

Pr [f (a) = Pℓ (a)] ≥ ε ,
a∈F2q ,ℓ∋a

for some constant ε > ε0 . Then, there is a bivariate degree d polynomial Q such that

Pr [f (a) = Q(a)] ≥ ε − ε0 .
a∈F2q

An added advantage of the above theorem is that it also works in the low-error regime. This
gives a single proof (in the bivariate setting) that works both for both the low-error and high-error
regimes. We can then bootstrap the bivariate low-error theorem using an argument similar to (but
simpler than) [FS95] to give an alternate proof of multivariate low-error Friedl-Sudan Theorem 1.2.
2
A set of parallel lines in any direction can be viewed as a set of lines in different directions through a single point
at infinity. In this sense, working with lines through a fixed point (a configuration that we refer to as a pencil), and
sets of parallel lines are essentially equivalent. For our arguments, the former happens to be a bit more natural.

7
Finally, we bootstrap the high-error version of the above theorem to yield the following multivariate
high-error result.
Theorem 1.5 (Multivariate low-degree test). There exists a constant τ ∈ (0, 1] such that for every
finite field Fq , ε0 ∈ (0, 1), m, and degree d satisfying ε0 > (d/q)τ , the following holds.
Suppose the points table f : Fm
q → Fq and degree-d lines oracle Pℓ satisfy

Pr [f (a) = Pℓ (a)] ≥ ε ,
a∈Fm
q ,ℓ∋a

for some constant ε > ε0 . Then, there is a bivariate degree d polynomial Q such that

Pr [f (a) = Q(a)] ≥ ε − ε0 .
a∈Fm
q

We thus not only prove the multivariate low-error LDT Theorem 1.5, but also given an alternate
(and arguably simpler) proof of the the multivariate low-error LDT Theorem 1.2. This yields a
completely self-contained treatment of the lines-point test in both error regimes.

1.3 Proof overview

We give a high-level overview of the main ideas in the proofs of our results. The overall structure
of our argument proceeds as follows — (1) Analyse the bivariate low-degree test, and (2) bootstrap
the argument to m-variate low-degree tests.
All of our key algebraic ideas already make an appearance in the proof of the bivariate low-degree
test, which is then lifted to the multivariate setting using combinatorial techniques, in particular
the expansion properties of the points-lines-planes affine Grassmannian.

1.3.1 Bivariate low-degree tests

Suppose are given a ‘points table’ f : F2q → Fq that passes the low-degree test with probability ε.
The argument for Theorem 1.4 proceeds as follows:

1. Find an ‘explainer’: We will show that there is a trivariate polynomial A(x, y, z) of (1, 1, d)-
weighted degree at most D that satisfies A(a, b, f (a, b)) = 0 for a structured set S of poly(ε)
fraction of the points (a, b) ∈ F2q .

2. Show that A has ‘low-degree roots’ on many lines: We then find a ‘good’ point (b1 , b2 ) ∈
F2q such that on many lines ℓb,u = {(b1 + tu1 , b2 + tu2 ) : t ∈ Fq } through this point, we have
a degree d polynomial Pb,u (t) such that A(b1 + tu1 , b2 + tu2 , Pb,u (t)) ≡ 0.

3. Show that A must have a global ‘low-degree root’: With the hypothesis from the
previous step, we use Lemma 1.3 to show that there must be a degree d polynomial Q(x, y)
such that A(x, y, Q(x, y)) ≡ 0.

8
 4. Show that f must have non-trivial agreement with Q(x, y) From the polynomial ob-
tained in the previous step, and the structure of S, we show that f and Q must agree on
Ω(ε4 · q 2 ) points.

Once we have non-trivial agreement as in Item 4, standard reductions in the low-degree testing
literature yields the stronger form as stated in Theorem 1.4.
Items 1 and 2 proceed along similar lines as in the results of Arora and Sudan [AS03], with some
additional care to ensure that the z-degree of A is function of just ε (and not d). With this additional
care, the argument is able to interpolate such an ‘explainer’ when the density of S is O(D/|F|) as
opposed to poly(D)/|F| in [AS03]. The key technical step is Item 3 which uses Lemma 1.3. This
lemma, in spirit, can be thought of as an analog of the celebrated lemma of Polischuk & Spielman
[PS94] in the high error setting. Item 4 is almost an immediate consequence of the previous steps.

Sketch of the proof of Lemma 1.3: Since we have a ‘local root’ for many lines, we first
observe that many of the polynomials Pb,u (t) must share the same constant term α. Using the fact
that (b1 , b2 ) is a ‘good’ point, this ensures some non-degeneracy properties for the point (b1 , b2 , α)
satisfying A(b1 , b2 , α) = 0. We then use Newton Iteration to construct an “approximate” root
Φk (x, y) such that A(b1 + x, b2 + y, Φk (x, y)) = 0 mod hx, yik+1 , for any choice of k. (That is,
Φk (x, y) is a root if we are willing to ignore some high-degree terms.) Finally, by using some
uniqueness properties guaranteed by Newton Iteration, we show that many of the Pb,u ’s must
infact be restrictions of Φd (x, y) to the respective lines. This allows us to eventually argue that
Φd (x, y) must infact be an global root of A(x, y, z).
In a broader sense, the above proof opens up the use of Hilbert’s Irreducibility Theorem in
[AS03] and makes appropriate changes to ensure the better dependence in parameters. A more
detailed discussion on the differences is provided below.

Dealing with fields of small positive characteristic: In general, proofs obtained via deriva-
tive based techniques like Taylor expansion or Newton iteration suffer from technical issues when
applied over fields of small characteristic. Among examples of such results are the results on poly-
nomial factorization [Kal89, KSS15], results on list decoding of multiplicity codes [GW13, Kop14,
Kop15, BHKS23] and hardness-randomness tradeoffs in algebraic complexity [GKSS22, And20]. In-
tuitively, the issue stems from the fact over a field of small characteristic, a polynomial can depend
on a variable, but its partial derivative with respect to this variable can still be identically zero, e.g.
if the polynomial is a function of xp over a finite field of characteristic p. Our outline also suffers
from these issues. However, we observe that these issues can be resolved and that the results extend
to all finite fields via one simple additional idea — while interpolating an explainer, ensure that
A(x, y, z) has a non-zero partial derivative with respect to z by just ignoring all monomials whose
exponent in z is divisible by p. As it turns out, this simple modification still allows us to proceed

9
with the rest of the argument without hardly any change to the parameters involved.

Technical differences with the analysis of Arora-Sudan: While our proof in the bivariate
case is conceptually similar to that of Arora-Sudan [AS03], there are some differences that lead to
better quantitative bounds. The first technical difference is that in the interpolation step of the
analysis, we interpolate a trivariate polynomial A(x, y, z) of (1, 1, d)-degree D, whereas Arora-Sudan
work with polynomials of total degree D. As a consequence, the z-degree of A can be bounded by
at most D/d and not just D. In particular, when D = O(d), the z-degree of A happens to be a
constant for us, whereas this is not the case in [AS03].
The second technical difference is in the statement of a lemma analogous to Lemma 1.3 that
Arora & Sudan prove. In their lemma (Lemma 39 in [AS03]), they construct a polynomial of degree
O(D 3 ) that essentially detects whether for u, v ∈ F, the restriction A(ut, vt, z) of A(x, y, z) has a
factor that is linear in z. More precisely, if A(x, y, z) does not have a factor that is linear in z,
then they construct a non-zero polynomial Q(x, y) of degree at most O(D 3 ) such that if Q(u, v) is
non-zero, then A(ut, vt, z) does not have a factor that is linear in z. Two points to note here are
that the degree of Q is polynomially larger than D, and that this statement talks about factors that
are linear in z, but is not sensitive to the degree of the such a factor in x, y (which could be as large
as D). As it turns out, the analysis of the bivariate low-degree test only ever cares about factors
that are linear in z and have degree at most d in x, y. Being aware of (x, y)-degree of these factors
of interest, and additionally considering polynomials of (1, 1, d)-degree at most D lets us construct
an analog of the polynomial Q above whose degree is just D. This reduction in degree of Q then
naturally translates to a reduction in the field size requirement in the overall analysis.

1.3.2 Bootstrapping to higher dimensions

Having proved the low-degree test theorem for dimension m = 2, we now need to bootstrap it to
higher dimensions. There are several possible ways to do this. One potential route is the following.
Given the lines-point bivariate LDT (i.e., m = 2), we obtain a degree-d planes oracle that passes
the planes-point test with non-trivial probability and we can then use the Raz-Safra bootstrapping
analysis [RS97] to bootstrap to arbitrary dimensions. This however causes ε0 (in Theorem 1.5) to be
at least Ω(m · (d/q)τ ). An alternate route to get around this linear dependence in m is the following
route. Bootstrap using the Raz-Safra analysis to dimension m = 3 and then use the Bhangale-
Dinur-Navon bootstrapping argument [BDN17] to bootstrap to arbitrary dimensions. The [BDN17]
τ
analysis (as written in their paper) requires ε0 to be at least Ω((d8/q) ). While this avoids the
m-dependence, it requires the field size q to be at least d8 . The [BDN17] analysis can be tighted
to yield a q = O(d) dependence by using the Friedl-Sudan analysis [FS95] instead of the Rubinfeld-
Sudan analysis [RS96] which they use. To avoid these issues and give a self-contained proof, we
give a direct bootstrapping argument from m = 2 to arbitrary dimensions (inspired of course by

10
bootstrapping analyses of [FS95, AS03, RS97, BDN17]). However, since we intend to lift from the
bivariate LDT (that is, the line-point test in a plane), some effort is required to make the broad
ideas of bootstrapping work in our regime. As mentioned earlier in the introduction, this direct
bootstrapping is inspired by the clean local-list-decoder for multivariate polynomial codes due to
Sudan, Trevisan & Vadhan [STV01].

As a warmup, let us first explain the bootstrapping for the low-error regime, a la Friedl-Sudan
[FS95]. In this setting, we are given ‘points table’ f : Fm
q → Fq and the best degree-d lines oracle
P (f,d) that fails the LDT with probability at most δ, we wish to show that there is a degree-d
polynomial Q(x1 , . . . , xm ) that is O(δ)-close to f . Following Friedl-Sudan, we define a corrected
(f,d)
function fcorr as follows: for any point y ∈ Fm
q , fcorr (y) is the most popular value of Pℓ (y)
(f,d) (f,d)
among all lines ℓ passing through y where Pℓ is the best-fit degree-d polynomial Pℓ agreeing
with f on the line ℓ (breaking ties arbitrarily). Let δf be the rejection probability of low-degree
(f,d)
test when run on the points table f and the best-fit lines oracle Pℓ for f . We will show that as
long as δf ≤ δ0 for some fixed constant (dependent on d/q ), the rejection probability of the corrected
function fcorr , names δfcorr is significantly smaller than that of f . More precisely, δfcorr ≤ δf /2. Once
we have shown this, we can repeat the self-correction procedure several times to eventually arrive
at a corrected function f∗ such that δf∗ = 0, in which case it is a low-degree polynomial. So, it
suffices to show that the corrected function fcorr passes the low-degree test with significantly better
probabibility than the original function f . To prove this, it suffices for us to show that for a random
(f,d) (f,d)
point x, the probability that two random lines ℓ, ℓ′ passing through x satisfy Pℓ (x) = Pℓ′ (x).
For any such triple (x, ℓ, ℓ′ ), consider the plane π containing ℓ and ℓ′ . If the function f restricted to
this plane π passes the low-degree with high probability, then the bivariate low-error LDT theorem
states that restricted to this plane, the function f , mostly behaves like a degree-d polynomial Qπ
(f,d) (f,d)
and both Pℓ and Pℓ′ are in fact restrictions of Qπ and hence equal to each other on the point
x. The the above arguments work only on average and not for every triple (x, ℓ, ℓ′ ). Nevertheless, we
show that this suffices to bootstrap to higher dimensions m and here we use the expansion properties
of the points-lines-planes affine Grassmannian. This proof is adapted from (and arguably simpler
than) the corresponding bootstrapping proof due to Friedl and Sudan [FS95], who use a different
high-dimensional expander consisting of points-lines-surfaces.
We now turn to bootstrapping in the high-error regime. Given a ‘points table’ f : Fm
q → Fq
that passes the LDT with probability ε, we wish to show that there is a degree d polynomial
Q(x1 , . . . , xm ) that agrees with f on poly(ε) places (proceeding from a ‘weak-agreement’ statement
to the statement in Theorem 1.5 again follows from standard reductions). As in the low-error case,
we would like to define a corrected function such that the corrected function passes the low-degree
test with significantly better probability than the original function. The primary issue here is
that there could be many candidate Q’s that have poly(ε) agreement and hence several different
choices for the corrected function. We use an additional advice (a random point x and the value of

11
the function f at the point x, namely f (x)) to disambiguate among the several different choices.
(x) (x)
One such corrected function gcorr : Fm → F is as follows. For any y ∈ Fm
q , gcorr (y) is the most
(x) (x)
popular value of Qπ (y) among all planes π containing both x and y such that Qπ is the unique
(x)
degree-d polynomial that “explains” the plane π and furthermore Qπ (x) = f (x). Why does such
a “explaining” polynomial Qπ exist? If the restriction of the function f to the plane π passes the
low-degree test with probability ε (which happens if the plane π is random), then the bivariate
LDT theorem states that there exists such a polynomial Qπ . This is precisely the bootstrapping
(x)
argument of Arora and Sudan [AS03]. However, then to show that this corrected function gcorr
passes the low-degree test with significantly better probability than f , one needs to consider a cube
and this is why the [AS03] bootstrapping argument required as base case both the m = 2 and m = 3
cases. To get around this dependence on the m = 3 case, we define an alternate correction function
(x)
fcorr (also disambiguated using the advice (x, f (x))), that we describe informally here:
(x)
Pick a random x ∈ Fm
q to use for constructing the correction. Define fcorr (x) = f (x).

For every other point y ∈ Fm

q , consider the line ℓx,y passing through x and y. Find the
list of all polynomials P (t) that agree with f on this line on at least poly(ε)-fraction of
(x)
places. Among those, if there is a unique P that takes value f (x) on x, set fcorr (y) =
P (y) and set it to ⊥ (or a random value in Fq ) otherwise.

In other words, the line joining x, y is being used to ‘correct’ the value at y but we are disambiguating
between the potential possibilities on this line by the value those polynomials take at the point x.
(x)
Having defined this corrected function fcorr , it is not hard to show for with non-trivial probability
(x)
over the choice of the advice point x, we have that the corrected function fcorr has non-trivial poly(ε)
agreement with the function f .
We now try to understand the behaviour of the corrected function on a random line ℓ. Consider
the plane π containing the advice point x and the line ℓ. Using the bivariate LDT theorem, we
can show that for a random plane π a non-trivial fraction of the points x in π, there exists a
(x) (x)
degree-d polynomial Qπ that has poly(ε)-agreement with f on π and Qπ (x) = f (x). We use
this coupled with the expansion properties of the lines-points-planes affine Grassmannian to show
(x)
that for a random x and a random line ℓ, the corrected function fcorr (using advice x) agrees with
(x)
the polynomial Qπ overwhelmingly on the line where π is the plane containing x and ℓ. This step
requires a careful analysis by using spectral properties of various natural subgraphs of the Affine
(x) (x)
Grassmanian. We thus, have, Prx,ℓ [δ(fcorr |ℓ , Qπ (x, ℓ)|ℓ ) ≤ γ] ≥ 1 − 2γ. This implies that for
(x)
a random x, the corrected function fcorr passes the low-degree test overwhelmingly (in fact with
probability 1 − 3γ). Hence, by Theorem 1.2, there exists a degree-d m-variate polynomial Q(x) such
(x) (x)
that δ(fcorr , Q(x) ) ≤ O(γ). Combining this with the fact that fcorr has poly(ε) agreement with the
function f , we obtain that f has poly(ε) − O(γ) = poly(ε) agreement with the polynomial Q(x)
provided γ ≪ poly(ε). This completes the overview of the bootstrapping argument.

12
1.4 Organisation of the paper
We begin with notation and preliminaries in Section 2 and then proof the main technical lemma
(Lemma 1.3) in Section 3. We then proceed to the analysis of the bivariate low-degree tests in
Section 4. Finally, we bootstrap the bivariate low-degree tests to m-variate low-degree tests in
Section 5.

2 Preliminaries
Notation:

1. For a polynomial f (x1 , . . . , xm ) and a vector a = (a1 , . . . , am ) ∈ Nm , we use the term e-

weighted degree to denote the largest value of a1 e1 + · · · + am em among monomials xe11 · · · xemm
in the support of f .
2. For two functions f, g : Fm → F, we define agree(f, g) = Pra∈Fm [f (a) = g(a)], the fraction of
points that the tables f and g agree on. Similarly, we will use δ(f, g) to denote Pra∈Fm [f (a) 6=
g(a)], the fraction of points that the tables f and g disagree on.
P
Lemma 2.1 (Averaging argument). Suppose x1 , . . . , xn ∈ [0, 1] such that xi ≥ µn. Then,

1. if S = {i ∈ [n] : xi ≥ µ/2}, then |S| ≥ µn/2, and

P
2. i∈S xi ≥ µn/2.

Proof. For Item 1, note that

X
µn ≤ xi ≤ µ/2 · (n − |S|) + |S| · 1 ≤ µn/2 + |S|
i
=⇒ µn/2 ≤ |S|.

For Item 2,
X X
xi ≥ µn − xi ≥ µn − (n − |S|) · µ/2
i∈S i∈S
/
≥ µn − (1 − µ/2) · µ/2 · n ≥ µn/2.

Polynomial identity lemma

Lemma 2.2 ([Ore22, DL78, Sch80, Zip79]). Let P (x) be a non-zero polynomial of total degree at
most D with coefficients over a field F and let S be an arbitrary subset of F. Then, the number of
zeroes of P on the product set S × S × · · · × S is at most D|S|m−1 .

13
Hasse Derivatives and Properties
Throughout the paper, we use the notion of Hasse derivatives of polynomials and some of their
basic properties. We start with the definition.
Definition 2.3. Let F be any field and A(x) ∈ F[x] be an m variate polynomial with coefficients in
F. Then, the Hasse derivative of A(x) with respect to a monomial xe , denoted by ∂¯xe (A), is defined
as the coefficient of the monomial ze when viewing A(s + z) as a polynomial in z variables with
coefficients in the ring F[x]. ♦
Hasse derivatives are an extremely useful and recurrent theme in computer science with a variety
of applications. For our proofs in the paper, we rely on the following properties.

Proposition 2.4. Let F be any field. Then, the following are true.

• For every d ∈ N, ∂¯ x (xd ) = dxd−1 .

• For every pair of polynomials A, B ∈ F[x], ∂¯(AB) = ∂¯(A) · B + B · ∂¯ (A).

• For any polynomial A ∈ F[x] of degree at least one in x, and a finite field F of characteristic
p, ∂¯ x (A) is identically zero if and only if there exists a polynomial à ∈ F[x] such that Ãp = A.

Discriminant
Definition 2.5 (Resultant and Discriminant). Let F be any field and A, B ∈ F[x] be non-zero
univariate polynomials in x of degree a, b respectively. Let A0 , A1 , . . . , Aa , B0 , B1 , . . . , Bb be elements
P P
of F such that A(x) = ai=0 Ai xi and B(x) = bj=0 Bj xj . Then, the Sylvester matrix of A and B
defined as follows.
 
A0 A1 ... Aa
 .. .. .. .. 
 . . . . 
 
 
 A0 A1 ... Aa 
 
Sylvester(A, B) = 
B0 . . . Bb 

 
 B0 ... Bb 
 
 .. .. .. 
 . . . 
B0 . . . Bb

And, the resultant of A and B is defined as the determinant of the Sylvester matrix of A and B.
Moreover, if B = ∂¯x (A), then the resultant of A and B is referred to as the Discriminant of A and
B and denoted by Discx (A, B). ♦

Lemma 2.6. Let F be any field and A, B ∈ F[x] be non-zero univariate polynomials in x. Then,
A, B have a non-trivial GCD (the degree of GCD is at least one in x) if and only if their resultant
is zero.

14
Definition 2.7. Let F be any field. A polynomial A(x) ∈ F[x] is said to be square-free if there does
not exist a polynomial B ∈ F of degree at least one such that B 2 divides A. ♦

Lemma 2.8. Let F be any field and A ∈ F[x] be a non-zero univariate polynomial such that ∂¯x (A)
is non-zero. Then, A is square-free if and only if the discriminant of A is non-zero.

2.1 Structure of minimal interpolating polynomials

In this section, we prove the following simple lemma that will be crucial to our analysis of the
bivariate test. The lemma essentially lets us assume some structural properties on the interpolating
polynomials without loss of generality, and these properties turn out to be important for our analysis,
especially when we are working over fields of small characteristic.

Lemma 2.9. Let F be any field and m ∈ N be an integer, and let S ⊆ Fm be a set of points. Consider
the space of polynomial QS,x1 = {Q(x1 , . . . , xm ) : Q(a) = 0 for all a ∈ S and ∂¯ x1 (Q) 6= 0}. For
any vector w ∈ Nm , the polynomial Q ∈ QS,x1 of lowest w-weighted degree satisfies Discx1 (Q) 6= 0
and is hence square-free.

Proof. Let Q ∈ QS,x1 be the polynomial of smallest w-weighted degree. On the contrary, let us
assume that Q is not square-free and say Q = P ℓ · R where P is irreducible with ℓ ≥ 2 and
gcd(P, R) = 1.
Consider the polynomial Q̃ = P · R. Clearly Q̃(a) = 0 for all a ∈ S as well, and if we can argue
that ∂¯x0 (Q̃) 6= 0, we have Q̃ ∈ QS,x1 thus yielding a contradiction to the minimality of Q.

∂¯ x1 (Q) = P ℓ · ∂¯ x1 (R) + ℓP ℓ−1 · R · ∂¯x0 (P ).

∂¯ x1 (Q̃) = P · ∂¯ x1 (R) + R · ∂¯x1 (P ).

Since we have ∂¯x1 (Q) 6= 0, we have that at least one of ∂¯x1 (P ), ∂¯ x1 (R) is non-zero. If exactly one of
∂¯x1 (P ), ∂¯ x1 (R) is non-zero, we immediately have that ∂¯ x1 (Q̃) 6= 0. Otherwise, since P is irreducible
and gcd(P, R) = 1, we have that P ∤ R · ∂¯x1 (P ). Therefore, P ∤ ∂¯x1 (Q̃) and this in particular forces
∂¯x1 (Q) 6= 0 yielding our required contradiction. Thus Q must have been square-free.

2.2 Power series roots

The following lemma is a standard application of Newton iteration in multivariate polynomial rings
and is an important ingredient of many of the multivariate factorization algorithms. We refer to
[Bür00] for a proof.

Lemma 2.10 (Newton iteration for roots). Let F be a field and let A(x, z) ∈ F[x, z] be an (m + 1)-
variate polynomial and let α ∈ F be a zero of multiplicity one of A(0, z), i.e., A(0, α) = 0 and
∂¯z (A(x, z))(0, α) = ∂¯z (A(0, z))(α) 6= 0. Then, for every k ≥ 1, there is a polynomial Φk (x) such
that the following are true.

15
 • A(x, Φk (x)) ≡ 0 mod hx1 , x2 , . . . , xm ik+1 ,

• α = Φk (0) ≡ Φk (x) mod hx1 , x2 , . . . , xm i,

• deg(Φk ) ≤ k.

Moreover, there is a unique polynomial that satisfies all these three properties.

3 Low-degree roots on restrictions to global low-degree roots

We prove the following technical lemma (a more formal statement of Lemma 1.3) that is the main
technical statement driving the analysis of our bivariate low-degree test. The lemma is true more
generally (in higher dimensions) and we state and prove it in that way.

Lemma 3.1. Let F be any field and let A(x, z) ∈ F[x, z] be a non-zero (m + 1)-variate polynomial
with z-degree at most dz , (1, 1, . . . , 1, d) degree at most D and let b ∈ Fm be a point such that the
univariate polynomial A(b, z) ∈ F[z] has no repeated roots. Let S ⊆ Fm be a set of directions such
that for every u ∈ S, there exists a univariate polynomial Pb,u (t) ∈ F[t] of degree at most d that
satisfies the identity

A(b + tu, Pb,u (t)) ≡ 0 .

If |S| > dz D|F|m−1 , then there exists a polynomial P (x) ∈ F[x] of total degree at most d such that

A(x, P (x)) ≡ 0 .

Moreover, there is a set S ′ ⊆ S of size at least |S|/dz such that for all u ∈ S ′ , P (x) when
restricted to the line in direction u through the point b equals Pb,u (t).

Proof. In its essence, the lemma above states that if for many different lines Lb,u through a point
b, the restriction A(Lb,u , z) ∈ F[t][z] of the polynomial A(x, z) has a low-degree polynomial root
Pb,u (t), then the original unrestricted polynomial A(x, z) ∈ F[x][z] has a low-degree root P (x).
Moreover, we are given that the point b is non-degenerate in the sense that the univariate polyno-
mial A(b, z) is square free. Qualitatively, the lemma is of flavor similar to the Hilbert’s irreducibility
theorem and the proof will be along similar lines but technically simpler and as it eventually turns
out, quantitatively better for our eventually applications.

For ease of notation, we will assume that b = 0 (by translating A if necessary) and use Pu to
denote Pb,u . Let hxi denote the ideal hx1 , . . . , xm i.
From the hypothesis of the lemma, we know that for every u ∈ S, there is a polynomial P0,u (t)
of degree at most d such that A(tu, Pu (t)) ≡ 0, which implies that A(0, Pu (0)) = 0. Therefore, if

16
{α1 , . . . , αℓ } is the multiset of roots over F for the polynomial A(0, z), we must have that Pu (0) = αi
for some i ∈ [ℓ]. As the deg(A(0, z)) ≤ dz , we have that ℓ ≤ dz . Hence, there must exist some i
such that αi = Pu (0) for at least D |F|m−1 many u ∈ S. Let us refer to this αi as just α, and let
R = {u ∈ S : Pu (0) = α}. Therefore, we have that Pu (0) = α for all u ∈ R, and also

A(tu, Pu (t)) = 0 =⇒ A(tu, Pu (t)) = 0 mod hxik+1 for all k ≥ 0.

On the other hand, since we are given that A(0, z) is square-free, the roots of this polynomial
are distinct and hence ∂¯ z (A(0, z)) 6= 0. Thus, by Lemma 2.10, there exist polynomials Φk (x) for
every k ≥ 0 such that

Φk (0) = α,
A(x, Φk (x)) = 0 mod hxik+1 .

By making the substitution xi 7→ txi , the above equation transforms to A(tx, Φk (tx)) = 0 mod tk .
In particular, we have

A(tu, Φk (tu)) = 0 mod tk for all u ∈ R.

Fix an arbitrary u ∈ R and consider the polynomial A′u (t, z) := A(tu, z) ∈ F[t, z]. We know that
A′u (0, α) = 0 is square-free and ∂¯ z (A′u (t, z))(0, α) 6= 0. Thus, by applying Lemma 2.10 on A′u (t, z)
to lift from α, there ought to be a unique polynomial Ψu (t) of degree at most d such that Ψu (0) = α
and

A′u (t, Ψu (t)) = A(tu, Ψu (t)) = 0 mod td+1 .

We already have two such candidates for Ψu (t), namely the polynomials Pu (t) and Φd (tu). Thus,
by the uniqueness asserted by Lemma 2.10, we have that

Pu (t) = Φd (tu) for all u ∈ R.

To finish the proof, we wish to argue that Φd (x) is a true root of A(x, z), i.e. it satis-
d+1
fies A(x, Φd (x)) = 0 (even without modhxi ). To this end, consider the polynomial B(x) =
A(x, Φd (x)). Since deg(Φd (x)) ≤ d and A has (1, . . . , 1, d)-weighted degree at most D, we have
deg(B(x)) ≤ D.
For every u ∈ R, we have Φd (tu) = Pu (t) and hence B(tu) = A(u, Φd (tu)) = A(tu, Pu (t)) = 0
as a polynomial in t, and thus B(u) = 0. This implies that the m-variate polynomial B of degree
at most D has more than D · |F|m−1 zeros in Fm . By the Polynomial Identity Lemma (Lemma 2.2),
conclude that B(x) = A(x, Φd (x)) is the zero polynomial.

17
4 The bivariate low-degree test
In this section, prove Theorem 1.4 which we recall for convenience.

Theorem 1.4 (Bivariate low-degree test). There exists a constant τ ∈ (0, 1] such that for every
finite field Fq , ε0 ∈ (0, 1) and degree d satisfying ε0 > (d/q)τ , the following holds.
Suppose the points table f : F2q → Fq and degree-d lines oracle Pℓ satisfy

Pr [f (a) = Pℓ (a)] ≥ ε ,
a∈F2q ,ℓ∋a

for some constant ε > ε0 . Then, there is a bivariate degree d polynomial Q such that

Pr [f (a) = Q(a)] ≥ ε − ε0 .
a∈F2q

Remark 4.1 (Version with a points table and a lines table). Many results in the low-degree testing
literature often consider a variant where we are given a ‘points table’ f : Fm → F, and a ‘lines table’
P : L(m) → F[x]≤d (where L(m) denotes the set of lines in Fm ) that assigns a polynomial of degree
at most d for every line. The version above is a special case where P is the canonical lines table
that assigns the best-fit degree d polynomial on each line. The discussion in the section carries over
in a straightforward manner to the more general setting of points and lines table, and we deal with
the special case purely to avoid some additional notational clutter. ♦
Before we proceed with the proof of the above theorem, we will fix some notation. Let f : F2 → F
be provided as an oracle. Throughout this section, we will be using LDTd to refer to the line-point
test:

• Pick x ∈R F2 and a random line ℓ ∋ x.

• Query the oracle on all points on the line ℓ and let Pℓ be the best-fit degree d polynomial.
• Accept if Pℓ (x) = f (x).

We will use εx to refer to Prℓ∋x [Pℓ (x) = f (x)].

To begin with, we will prove the following “weaker” statement for the low-degree test.

Theorem 4.2. Suppose ε > 0, and d is a positive integer. Let F is a finite field with q elements
q > C · d/ε7 for a large enough constant C. Suppose f : F2 → F that passes LDTd with probability at
least ε, that is

Pr [f (a) = Pℓ (a)] ≥ ε.
a,ℓ∋a

18
Then, there is a polynomial Q(x, y) of degree at most d such that

Pr [f (a) = Q(a)] ≥ Ω(ε4 ).

a∈F2

Although the above theorem appears to yield a weaker agreement than claimed by Theorem 1.4,
it would turn out that Theorem 4.2 yields Theorem 1.4 via standard reductions that we elaborate
on towards the end of this section. Thus, we proceed towards proving Theorem 4.2.

As a first step, we show the existence of a low-degree trivariate polynomial that explains that
the given function passes the line point test with a constant probability.

Theorem 4.3 (Interpolation for LDT). There exist constant c1 , c2 ∈ N such that for every ε ∈ (0, 1],
any d ∈ N and finite field F of size q and characteristic p with q > c1 · d/εc2 , the following is true.
Let f : F2 → F be an oracle that passes LDTd with probability ε. Then, there is a non-zero
polynomial A(x, y, z) and a subset S ⊆ F2 such that

• deg1,1,d (A) = O(d/ε2 ) and |S| = Ω(ε2 q 2 ),

P 3 2
• x∈S εx = Ω(ε · q ),

• for every (a, b) ∈ S, we have ε(a,b) = Prℓ∋(a,b) [Pℓ (a, b) = f (a, b)] ≥ ε/2,
• for every (a, b) ∈ S, we have A(a, b, f (a, b)) = 0.
• ∂¯ z (A) and Discz (A) are not identically zero.

The proof is mostly along the lines of Arora and Sudan [AS03] but with a tighter analysis
and some care for fields of small characteristic. We present it in its entirety, in Appendix A, for
completeness and to make the changes clearer to follow.

We now use Theorem 4.3 to complete the proof of Theorem 4.2.

Proof of Theorem 4.2. Suppose f : F2 → F is given and we know that f passes LDTd with proba-
bility at least ε. By Theorem 4.3, we have a set S ⊆ F2 with |S| = γ · q 2 = Ω(ε2 · q 2 ) and a non-zero
polynomial A(x, y, z) with deg1,1,d (A) = D = O(d/ε2 ) such that for all (a, b) ∈ S we have

• A(a, b, f (a, b)) = 0,

• Prℓ∋(a,b) [Pℓ (a, b) = f (a, b)] ≥ ε/2.

We also know that ∂¯z (A) and Discz (A) are both non-zero polynomials. We would like to find an
(a, b) ∈ S such that the following guarantees hold:

(i) ΓA (a, b) 6= 0 where ΓA (x, y) = discz (A) = Resz (A, ∂z A),

19
 (ii) There are Ω(ε2 · (q 2 − 1)) choices for (α, β) ∈ F2 \ {(0, 0)}, such that there is a degree d
polynomial Pα,β (t) such that

A(a + αt, b + βt, Pα,β (t)) = 0

with Pα,β (0) = f (a, b), and Pα,β (a + αt, b + βt) = f (a + αt, b + βt) for at least Ω(ε2 · q) many
t ∈ F.

If we can find such an (a, b), then Lemma 3.1 would imply that there is a polynomial Q(x, y) of degree
at most d such that A(x, y, Q(x, y)) = 0 with Q|ℓ = Pα,β (t) where ℓ is the line {(a + αt, b + βt) : t ∈ F}
for any (α, β) satisfying Item (ii). Thus, for any (a′ , b′ ) = (a + αt, b + βt) that satisfy Pα,β (t) =
f (a′ , b′ ), we have Q(a′ , b′ ) = Pα,β (t) = f (a′ , b′ ). Thus, Q agrees with f on Ω(ε4 · q 2 ) locations, as
claimed by the theorem.

We now work towards guaranteeing Items (i) and (ii). Note that we have q 2 + q lines in F2 ,
and each point x ∈ F2 has q + 1 lines containing it, and each line has q points on it. Consider
the following bipartite graph where the set of left vertices is F2 and the set of right vertices are
all lines in F2 . We connect a point x ∈ F2 on the left to a line ℓ on the right if x ∈ S, x ∈ ℓ
and Pℓ (x) = f (x). Note that every x ∈ S has εx · (q + 1) edges in this bipartite graph. If µ > 0
P
such that µq 2 = 3 2
x∈S εx = Ω(ε · q ) (by Theorem 4.3), the number of edges in this graph is
P
(q + 1) · x∈S εx = µ · (q 3 + q 2 ).
For a line ℓ, let yℓ = deg(ℓ)/q where deg(ℓ) refers to the degree in this graph. Since each σℓ ∈ [0, 1]
P
and σℓ = µ(q 2 + q). By Lemma 2.1, if L = {ℓ : σℓ > µ/2} (the ‘high-degree lines’), then at least
µ/2 · (q 3 + q 2 ) edges of the graph are incident on lines in L — let us refer to these edges as “marked
edges”. Again by Lemma 2.1, there are at least µ/4 ·q 2 vertices in the left that have at least µ/4 ·(q +1)
“marked edges” incident on it — let us call this set of vertices S ′ ⊆ S.
Thus we now have a set S ′ ⊆ S ⊆ F2 with |S ′ | ≥ µ/4 · q 2 = Ω(ε3 · q 2 ) such that each x ∈ S ′ such
that at least µ/4-fraction of lines ℓ through it satisfy the following two properties:

• Pℓ (x) = f (x),
• the line ℓ contains at least µ · q/2 points of S.

Guaranteeing Item (i): The polynomial A(x, y, z) has deg1,1,d ≤ D = O(d/ε2 ) and hence we have
dz = degz (A) = O(1/ε2 ). Therefore, the z-discriminant of A, namely Discz (A) is the determinant
of a O(dz ) × O(dz ) matrix, each of whose entries is a polynomial in x, y of degree at most D. We
recall from the last item of Theorem 4.3 that Discz (A) is not identically zero.
Thus, deg(Discz (A)) = O(D · dz ) = O(d/ε4 ). By Lemma 2.2, there are at most O(dq/ε4 ) points
in F2 that the polynomial Γ vanishes on. Since |S ′ | = Ω(ε3 · q 2 ) and q > C · d/ε7 for a large enough
constant C, there must be some (a, b) ∈ S ′ that does not make Γ zero.

20
Guaranteeing Item (ii): Let (α, β) ∈ F2 \ (0, 0) such that ℓ = {(a + αt, b + βt) : t ∈ F} is one
of the ‘high-degree’ lines in L containing (a, b). Let Pℓ be the best-fit degree d polynomial on this
line. Since (a, b) ∈ S, we have Pℓ (0) = f (a, b).
Let the polynomial B(t) be defined as B(t) := A(a + αt, b + βt, Pℓ (t)). Clearly, B has degree at
most D. For any t ∈ F such that (a′ , b′ ) = (a + αt, b + βt) ∈ S, we have Pℓ (t) = f (a′ , b′ ) and hence
B(t) = A(a′ , b′ , f (a′ , b′ )) = 0. Since there are at least µ · q/2 = Ω(ε2 · q) such points, the polynomial
B(t) has Ω(ε2 · q) roots but has degree at most D = O(d/ε2 ). Since q > C · d/ε7 , we must have that
B(t) = A(a + αt, b + βt, Pℓ (t)) is identically zero.
As (a, b) is adjacent to µ/4 · (q + 1) lines, there are at least µ/4 · (q + 1) · (q − 1) = µ/4 · (q 2 − 1)
choices for (α, β) ∈ F2 \ (0, 0) such that there is some degree d polynomial Pα,β such that

A(a + αt, b + βt, Pα,β (t)) = 0.

Thus, we can now invoke Lemma 3.1 to get the desired conclusion. This completes the proof of
Theorem 4.2.

4.1 Some consequences of Theorem 4.2

Given Theorem 4.2, we conclude this section with a few variants that will be useful for us in the
proof of the high dimensional case. These variants, despite being seemingly stronger in appearance
follows immediately from Theorem 4.2 via standard techniques in the low-degree testing literature.
For the sake of completeness, we include formal statements and proofs in Appendix B.

Theorem 4.4 (List-decoding version of the bivariate LDT). For every ε0 ∈ (0, 1), finite field Fq
1/16
and degree d satisfying ε0 > Ω((d/q) ), the following holds:

For every f : F2q → Fq and every ε > ε0 , there is a (possibly empty) set {Q1 , . . . , Qt } of
polynomials of degree at most d such that t ≤ 2/ε8 and agree(f, Qi ) ≥ ε8 for all i ∈ [t]
and

Pr [a is ε-good and f (x) ∈

/ {Q1 (x), . . . , Qt (x)}] ≤ ε0 .
a∈F2q

The other is the ‘high-agreement variant’, namely Theorem 1.4:

Theorem 1.4 (Bivariate low-degree test). There exists a constant τ ∈ (0, 1] such that for every
finite field Fq , ε0 ∈ (0, 1) and degree d satisfying ε0 > (d/q)τ , the following holds.
Suppose the points table f : F2q → Fq and degree-d lines oracle Pℓ satisfy

Pr [f (a) = Pℓ (a)] ≥ ε ,
a∈F2q ,ℓ∋a

21
for some constant ε > ε0 . Then, there is a bivariate degree d polynomial Q such that

Pr [f (a) = Q(a)] ≥ ε − ε0 .
a∈F2q

Proofs of Theorems 1.4 and 4.4. Both the above theorems are immediate consequences of a generic
reduction from the ‘weak form of LDT’ to the ‘list-decoding form of LDT’ (Lemma B.4) and the
‘high-agreement form of LDT’ (Lemma B.6) used along with Theorem 4.2.

Theorem 1.4 is also of interest in the low error regime, where ε is close to 1. In this case, Theo-
rem 4.2 only lets us conclude a constant fraction agreement between f and a low-degree polynomial,
whereas the Theorem 1.4 asserts that f agrees with a low-degree polynomial on almost all inputs.
In the low error regime, in order to lift the analysis from the bivariate case to the multivariate case
for our proof, we rely on the following theorem, which, as we show in the proof also follows from
Theorem 1.4.
To avoid confusion with the low-error regime, we will use δ to denote the probability that f fails
the low-degree test.
Theorem 4.5 (Low-error high-agreement bivariate LDT). There is a large enough constant C such
that for any finite field Fq and degree d satisfying q > C · d, and any δ < 0.01 the following holds:
Suppose f : F2q → Fq passes the LDTd with probability 1 − δ, then there is a bivariate
degree d polyomial Q such that

Pr [f (x) = Q(x)] ≥ 1 − 2δ.

x∈F2q

Proof. Since f passes LDTd with probabaility at least (1 − δ), we have from Theorem 1.4 that there
exists a degree d polynomial Q such that

Pr [f (a) = Q(a)] ≥ 1 − δ − ε0 .
a∈F2

 
1
where ε0 = Ω (d/q) /20 . We note that this agreement is weaker than the conclusion of the above
theorem when δ is smaller than ε0 . Let C be chosen large enough so that δ + ε0 ≤ 0.01 + ε0 <
1/20 · (1 − d/q).
An important ingredient of the proof of this theorem is the definition of a corrected version of f
that we denote by fcorr , which is defined as follows.
n o
(f,d)
fcorr (a) := pluralityℓ∋a Pℓ (a) .

(f,d)
where ℓ is a line through a and Pℓ denotes a univariate polynomial of degree d that is closest to
the restriction of f on ℓ.

22
 The theorem is an immediate consequence of the following two claims

Claim 4.6. Pra∈F2 [f (a) = fcorr (a)] ≥ 1 − 2δ.

Claim 4.7. For all a ∈ F2 , we have fcorr (a) = Q(a).

Proof of Claim 4.6. Let B denote the set of bad points in F2 defined as follows
 
2 (f,d)
B := a ∈ F : Pr [Pℓ (a) 6= f (a)] ≥ /2 .
1
ℓ∋a

Clearly, if a ∈
/ B, then we have fcorr (a) = f (a). Hence, we have Pra [f (a) 6= fcorr (a)] ≤ Pra [a ∈ B].
On the other hand,

δ ≥ Pr[f fails LDTd on (x, ℓ)]

x,ℓ
(f,d)
≥ Pr[a ∈ B] · Pr [Pℓ (a) 6= f (a) | a ∈ B]
a∈F2
ℓ∋a

≥ Pr [f (a) 6= fcorr (a)] · 1/2

a∈F2

which implies that Pra [f (a) 6= fcorr (a)] ≤ 2δ and hence Pra [f (a) = fcorr (a)] ≥ 1 − 2δ.

Proof of Claim 4.7. As mentioned earlier, we know that f has fractional agreement of at least
(1 − δ − ε0 ) with the degree d polynomial Q. Let B be the set of inputs where f and Q disagree.
Clearly, |B| < q 2 (δ + ε0 ). Let a ∈ F2 be an arbitrary input. Since lines through a are a uniform
cover of the space F2 \ {a}, we have that for all large enough q,
 
q 2 (δ + ε0 )
E [|ℓ ∩ B|] ≤ q · ≤ 2q(δ + ε0 ).
ℓ∋a q2 − 1

Thus, by Markov’s inequality, Prℓ∋a [|ℓ ∩ B| > 8q(η + ε)] < 1/4.
Now, let ℓ be a line through a such that |ℓ ∩ B| < 8q(δ + ε0 ). For every such line ℓ, the
disagreement between Q and f on ℓ is at most 8q(δ + ε0 ) + 1 ≤ 10q(δ + ε0 ) < 1/2 · (1 − d/q). Since
this is less than half the minimum distance of a Reed-Solomon code of degree d and block-length q,
(f,d)
we have that Pℓ must equal Q |ℓ .
Hence, we have that on 3/4-th of the lines ℓ through a, the restriction Q |ℓ equals the polynomial
(f,d)
Pℓ . From the definition of fcorr , it now follows that fcorr (a) = Q(a). Since a is an arbitrary point
in F2 , we have that the fcorr and Q agree everywhere on F2 .

That concludes the proof of Theorem 4.5.

23
5 Lifting to m-variate low-degree tests
In this section, we bootstrap the low-degree test from 2 dimensions to m dimensions. This boot-
strapping is inspired and adapted from the corresponding bootstrapping arguments due to Friedl-
Sudan [FS95], Arora-Sudan [AS03] and Bhangale-Dinur-Navon [BDN17] and the local-list-decoder
for multivariate polynomial codes due to Sudan, Trevisan & Vadhan [STV01]. The bootstrapping
arguments heavily use the expansion properties of the points-lines-planes affine Grassmannian. So,
we first mention some preliminaries that we would need..

5.1 Preliminaries
Theorem 5.1 ((Weak) Johnson Bound). For any function f : F → F, degree parameter d and
ε ∈ (0, 1), let P1 , . . . , Pr be the set of all degree-d polynomials that have agreement at least ε with f
(i.e, Prx∈F [f (x) = P (x)] ≥ ε). We say that a point x ∈ F is non-unique with respect to function f ,
degree d and agreement ε, if there exist two distinct polynomials Pk 6= Pk′ , 1 ≤ k, k′ ≤ r such that
Pk (x) = Pk′ (x) or in short, “x ∈ nonuniquedε (f )”.
p
If ε ≥ 2 d/q, then the following bounds hold.

1. r ≤ 2/ε.
r


2. The number of points in nonuniquedε (f ) is at most 2 · d ≤ 2d/ε2 .

Definition 5.2. Let G = (A, B, E) be a bi-regular bipartite graph, and let M ∈ RA×B be the
adjacency matrix normalized such that kM 1k = 1, denote by λ(G) the value
 
kM vk
λ(G) = max .
v⊥1 kvk

We will refer to λ(G) as the second eigen-value of G. ♦

This is really the second largest singular value of M , with a different normalization (such that
the maximal singular value equals 1).
The following is the classic Expander Mixing Lemma (for bipartite graphs).

Lemma 5.3 (Expander Mixing Lemma). Let G = (A, B, E) be a biregular bipartite graph with
second eigen-value λ. Then for any two functions g : A → R and h : B → R with means and
variances µg , µh and σg2 and σh2 , we have

Pr [g(a) · h(b)] − µg · µh ≤ λ · σg · σh
(a,b)∈E

We will be using the following corollary of the above Expander Mixing Lemma.

24
Lemma 5.4 (Corollary of Expander Mixing Lemma [BDN17, Lemma 6]). Let G = (A, B, E) be a
biregular bipartite graph with second eigen-value λ. Then for any set A′ ⊆ A of measure µ and any
E ′ ⊆ E, we have

Pr [(a, b) ∈ E ′ ] − Pr [(a, b) ∈ E ′ ] ≤ λ/√µ.

a∼A′ ,b∼NG (a) b∈B,a∼NG (b)∩A′

The following are well-known properties of the points-lines-planes affine Grassmannian in Fm

q .

Theorem 5.5. The second eigen-value of the following biparite inclusion graphs is as follows:

1. For all m ≥ 2 and G = G(Fm , L (m) ), λ(G) = 1/√q.

2. For all m ≥ 2, G = G(Fm , P (m) ), λ(G) = 1/q.

3. For all m ≥ 2, G = G(L (m) , P (m) ), λ(G) = √1

q · (1 + o(1/√q)).

(m) (m) √
4. For all m ≥ 2 and G = G(Lx , Px ), λ(G) = 1/ q+1.

5.2 The bootstrapping statements

(f,d)
Given a function f : Fm → F and a line ℓ in Fm , we let Pℓ be the best fit degree-d univariate
polynomial (presented as a list of evaluations) that agrees with f on the line ℓ (if there is more
than one such polynomial, we break ties arbitrarily). We now, define the following quantities for
any function f : Fm → F, line ℓ and plane π:

(f,d)
δf (ℓ) := Pr [Pℓ (x) 6= f (x)],
x∈ℓ
(f,d)
δf (π) := E [δf (ℓ)] = Pr [Pℓ (x) 6= f (x)],
ℓ∈π ℓ∈π
x∈ℓ
(f,d)
δf := E[δf (π)] = Prm [Pℓ (x) 6= f (x)]. (5.6)
π ℓ∈F
x∈ℓ

Bootstrapping in the low-error regime

In Section 5.3, we bootstrap the low-error bivarariate LDT Theorem 4.5 to prove Theorem 1.2,
a similar theorem for m dimensions. This bootstrapping argument is an adaptation of a similar
argument due to Friedl and Sudan [FS95], the main difference being that we use the expansion
properties of the points-lines-planes affine Grassmannian rather than that of a more complicated
points-lines-surfaces HDX used in [FS95].

Theorem 1.2 (Multivariate low-error LDT [FS95]). There is a constant C large enough such that
for every finite field Fq , m > 0, degree d satisfying q > C · d, the following holds:

25
 Suppose the points table f : Fm
q → Fq and the degree-d lines oracle Pℓ satisfy

Pr[f (a) 6= Pℓ (a)] ≤ δ,

a,ℓ

for some 0 < δ < 0.01, then there exists a m-variate degree-d polynomial Q such that
δ(f, Q) ≤ 4δ.

Bootstrapping in the high-error regime

In the subsequent Section 5.4, we perform a similar bootstrapping for the high-error regime to
obtain the following theorem from bivariate low-degree test (Theorem 4.2). This bootstrapping
argument is inspired from the corresponding arguments due to Arora-Sudan [AS03] and Bhangale-
Dinur-Navon [BDN17]. The argument presented here, while elementary, requires a careful analysis
using the spectral properties (repeated applications of Lemmas 5.3 and 5.4) of various subgraphs
of the points-lines-planes affine Grassmannian. The key improvement from [AS03] and [BDN17] is
that we bootstrap from the base case of m = 2 while previous arguments worked with a base case
of at least m ≥ 3.

Theorem 5.7 (high-error regime m-variate LDT). For every ε0 ∈ (0, 1), and finite field Fq and
1/48
degree d satisfying ε0 > Ω((d/q) ), the following holds.

If the points table f : Fm → F and degree-d lines oracle ℓ 7→ Pℓ satisfy

Pr [f (x) = Pℓ (x)] ≥ 5ε,

x,ℓ∋x

then there exists an m-variate degree-d polynomial Q such that Prx [f (x) = Q(x)] ≥ ε2 .

As in the bivariate case, the above ‘weak form of the low-degree test’ (Theorem 5.7) can be
reduced to the ‘high-agreement form of the low-degree test’ (Theorem 1.5) using standard transfor-
mations (Lemma B.6).

5.3 Low-error regime: Proof of Theorem 1.2

Given a function f : Fm → F, we define the self-correction fcorr of the function f as follows:
n o
(f,d)
fcorr (x) := pluralityℓ∋x Pℓ (x) .

Let ε0 , δ0 ∈ (0, 1), the field Fq and degree d satisfy ε0 ≥ Ω((d/q) /20 ) as in the hypothesis of .
1

Define δ0 := 1/20(1 − d/q) − ε0 . By the hypothesis of Theorem 1.2, we have that δf ≤ δ0/2.
The theorem follows from the two claims.

Claim 5.8. δ(f, fcorr ) ≤ 2δf .

26
Claim 5.9. If q > 100/δ02 and δf ≤ δ0/2, then δfcorr ≤ δf/2.
Proof of Theorem 1.2. Define a sequence of functions f (0) , f (1) , . . . : Fm → F as follows: f (0) := f


and for i ≥ 1, f (i) := f (i−1) corr . We then have, from Claim 5.9, that δi := δf (i) ≤ δf/2i . For
any fixed q and m, since δf can only take a set of finitely many values, we have that δi eventually
becomes 0, i.e., there exists i∗ such that δi∗ = δf (i∗ ) = 0. Then, f (i∗ ) is a degree-d m-variate
polynomial, say Q. Now, by Claim 5.8, we have

  Xi∗   Xi∗ i∗
X
δ(f, Q) = δ f (0) , f (i∗ ) ≤ δ f (i−1) , f (i) ≤ 2δf (i−1) ≤ 2 δf/2i−1 ≤ 4δf .
i=1 i=1 i=1

This completes the proof of Theorem 1.2 (assuming the two claims).

We now proceed to prove the two claims.

(f,d)
Proof of Claim 5.8. Given the points table f : Fm → F and lines oracle ℓ 7→ Pℓ , we define the set
(f,d)
of “bad” points as follows: BAD := {x ∈ Fm : Prℓ∋x [Pℓ (x) 6= f (x)] ≥ 1/2}. Clearly, if x ∈
/ BAD,
we have fcorr (x) = f (x). Hence, δ(f, fcorr ) ≤ Pr[x ∈ BAD]. On the other hand, we have

(f,d)
δf = Prm [Pℓ (x) 6= f (x)]
x∈F
ℓ∋x
(f,d)
≥ Pr[x ∈ BAD] · Prm [Pℓ (x) 6= f (x) | x ∈ BAD]
x∈F
ℓ∋x

≥ δ(f, fcorr ) · 1/2.

Hence, δ(f, fcorr ) ≤ 2δf .

5.3.1 Self-correction passes LDT with better probability (Proof of Claim 5.9)

This is the heart of the bootstrapping argument, where we use the (high-dimensional) expansion of
the points-lines-planes Affine Grassmannian complex to show that the self-corrected function fcorr
passes the LDT with even better probability than the original function f .
We begin by showing the following bound on δfcorr .

(fcorr ,d)
δfcorr = Pr [Pℓ (x) 6= fcorr (x)]
x,ℓ∋x
(f,d)
≤ Pr [Pℓ (x) 6= fcorr (x)]
x,ℓ∋x
(f,d) (f,d)
≤ Pr
x
[Pℓ (x) 6= Pℓ′ (x)].
ℓ,ℓ′ ∋x

(f,d)
The first inequality follows since Pℓ cannot perform any better than the best-fit degree-d poly-
(f ,d)
nomial Pℓ corr . The second inequality follows since for each x ∈ Fm , fcorr (x) is the most popular

27
 (f,d)
value among Pℓ (x) as ℓ varies over all lines ℓ through x and hence the probability (over ℓ) that
(f,d)
fcorr (x) = Pℓ (x) is at least the collision probability that for two independently chosen lines ℓ, ℓ′
(f,d) (f,d)
through x, we have Pℓ (x) = Pℓ′ (x).
(f,d) (f,d)
It thus suffices to bound the probability that Pℓ (x) = Pℓ′ (x) where x, ℓ, ℓ′ are chosen as
follows: x is picked uniformly from Fm , ℓ, ℓ′ are independently chosen to be two lines through x
in Fm . An equivalent way of picking this triple is first picking a random plane π in Fm , a point x
in the plane π and two independent lines ℓ, ℓ′ in the plane π that contain x. We now define three
(bad) events E1, E2 and E3 (based on the choice of π, ℓ, ℓ′ ) such that (1) if none of the three events
(f,d) (f,d)
happen, then Pℓ (x) = Pℓ′ (x) and (2) the probability of each event is at most δf /6. This will
complete the proof of the claim.

Event E1(π): δf (π) ≥ δ0 .

Consider the bipartite lines-planes incidence graph G(L (m) , P m ) in Fm which has second
eigen-value at most 2/√q (by Theorem 5.5-3). We know that δf (π) = Eℓ∈π [δf (ℓ)] and δf =
Eℓ [δf (ℓ)]. Consider the functions g : L (m) → R and h : P (m) → R defined as follows: g(ℓ) :=
δf (ℓ) and h(π) := 1[δf (π) ≥ δ0 ]. These functions satisfy µg = δf , µh = Pr[E1] =: µ,
σg2 = Varℓ [δf (ℓ)] ≤ δf and σh2 = Varπ [1[δf (π) ≥ δ0 ]] ≤ µ. Applying Lemma 5.3 to the graph
G with functions g and h as defined above, we have
p
µ · δ0 − µ · δf ≤ λ · µ · δf .

Equivalently, µ ≤ (λ/(δ0 −δf ))2 · δf . Choosing q ≥ 100/δ2 ,

0 we have µ ≤ δf/6 (since λ ≈ 1/√q and
δf ≤ δ0 /2).
Furthermore, if event E1 does not happen then by Theorem 4.5 (the bivariate LDT in the
low-error regime), we have that there exists a bivariate degree-d polynomial Qπ on the plane
π such that δ(f |π , Qπ ) ≤ 2δf (π) ≤ 2δ0 .

Event E2(π, ℓ): ¬E1 and δ(f |ℓ , Qπ |ℓ ) ≥ 4δ0 . (Here Qπ is the bivariate degree-d polynomial that is
guaranteed to exist since E1 does not occur.)
To begin with let us fix a plane π such that E1 does not occur. We will later randomize over
the choice of the plane. Since E1 does not occur, we have there exists a bivariate degree-d
polynomial such that δ(f |π , Qπ ) ≤ 2δf (π) ≤ 2δ0 .
Consider the bipartite points-line incidence graph G(π, L (π) ) in the plane π which has second
eigen-value at most 1/√q (by Theorem 5.5-1). Let BADπ be the set of lines ℓ in π such that
δ(f |ℓ , Qπ |ℓ ) ≥ 4δ0 . Consider the functions g : π → R and h : L (π) → R defined as follows:
g(x) := 1[f (x) = Qπ (x)] and h(ℓ) := 1[ℓ ∈ BADπ ]. These functions satisfy µg = δ(f |π , Qπ ),
µh = µ(BADπ ) =: µπ , σg2 = Varx [[f (x) 6= Qπ (x)]] ≤ µg = δ(f |π , Qπ ) and σh2 = Varℓ [1[ℓ ∈
BADπ ]] ≤ µπ . Applying Lemma 5.3 to the graph G with functions g and h as defined above,

28
 we have
p
µπ · 4δ0 − µ · δ(f |π , Qπ ) ≤ λ · µπ · δ(f |π , Qπ ).

Equivalently, µπ ≤ (λ/(4δ0 −δ(f |π ,Qπ )))2 ·δ(f |π , Qπ ) ≤ (λ2/2δ02 )·δf (π) (since δ(f |π , Qπ ) ≤ 2δf (π) ≤
2δ0 ). Choosing q ≥ 3/δ02 , we have that for this choice of π, µπ ≤ δf (π)/6 (since λ ≈ 1/√q and
δf ≤ δ0 /2).
We now average over π as follows:

Pr[E2] = E [¬E1(π) · µπ ] ≤ E [¬E1(π) · (δf (π)/6)] ≤ E[δf (π)/6] = δf/6.

π,ℓ π π π

Observe that if for a particular choice of random π and ℓ, events E1 and E2 do not occur,
then there exists a bivariate degree-d polynomial Qπ such that δ(f |π , Qπ ) ≤ 2δf (π) ≤ 2δ0 and
furthermore δ(f |ℓ , Qπ |ℓ ) ≤ 4δ0 . If 1 − 8δ0 ≥ d/q, then Qπ |ℓ is the (unique) best-fit degree-d
(f,d)
polynomial to f |ℓ , i.e., Pℓ = Qπ |ℓ .

Event E3(π, ℓ′ ): ¬E1 and δ(f |ℓ′ , Qπ |ℓ′ ) ≥ 4δ0 .

This event is identical to E2 and hence Prπ,ℓ′ [E3] ≤ δf /6.

Clearly, if events E1, E2 and E3 do not occur (for a particular choice of π, ℓ, ℓ′ and x), we have
(f,d) (f,d)
that there exists a degree-d bivariate polynomial Qπ such that Pℓ = Qπ |ℓ and Pℓ′ = Qπ |ℓ′ .
(f,d) (f,d)
Hence, Pℓ (x) = Q(x) = Pℓ′ (x). This completes the proof of the claim.

5.4 High-error regime: Proof of Theorem 5.7

A degree-d lines oracle assigns to each line in Fm , a a degree-d polynomials (presented as a list of
evaluations, i.e., a Reed-Solomom codeword) or ⊥. We say that the degree-d lines oracle is ε-well-
behaved with respect to the function f : Fm → F if for every line ℓ, Pℓ is well-defined (i.e, 6= ⊥) if
and only if Prx∈ℓ [f (x) = Pℓ (x)] ≥ ε.
We say a point x ∈ Fm is ε-good if it agrees with at least an ε-fraction of lines that pass through
it, i.e.,

Pr [f (x) = Pℓ (x)] ≥ ε.
ℓ∋x

We will refer to the set of ε-good points in Fm as ε-GOOD. We will need the above notation, both
when the ambient dimension is m = 2 (i.e., a plane) or general m (i.e., Fm ). To distinguish these
two cases, in the former we say “x is ε-good with respect to plane π” , while in the latter we just
say “x is ε-good”. Sometimes, we also say “x is ε-locally-good” vs. “x is ε-globally-good”.
We say that “x is ε-explained with respect to a plane π” if there exists an bivariate degree-d
polynomial Q (defined on the plane π) such that (1) Prz∈π [f (z) = Q(z)] ≥ ε and (2) f (x) = Q(x).

29
 We will assume the the bivariate LDT, given by Theorem 4.4, which states the following (rewr-
riten in the language of “”ε-explained”-ness). Let ε0 := Ω((d/q)1/16 ). For every ε ≥ ε0 , the following
holds. For any function f : F2 → F,

Pr[x is ε-good but not ε8 -explained ] ≤ ε0 . (5.10)

x

We prove Theorem 5.7 by defining a self-corrected function as in the low-error regime and
showing that the corrected function passes the LDT with significantly higher probability. However,
unlike the low-error regime, there are several candidate corrected functions and we disambiguate
among them using an additional advice (x, σ) ∈ Fm × F.
x,σ
For x ∈ Fm and σ ∈ F, define fcorrδ : F
m → (F ∪ ⊥) as follows: For any y ∈ Fm \ {x}, let P be

the unique degree-d univariate polynomial on the unique line ℓ = ℓx,y through x and y such that
(1) Prz∈ℓ [f (z) = P (z)] ≥ δ and (2) P (x) = f (x). If there is no such polynomial P or there is more
x,σ x,σ
than one such polynomial, we set P := ⊥. Finally, we set fcorrδ (y) := P (y). Also, set fcorrδ (x) := σ.
x x,f (x) x
Finally we define fcorrδ
:= fcorrδ . We will refer to the function fcorrδ
as the x-corrected function.
x
The following lemma states for a random ε-good x, the corrected function fcorr passes the LDT
δ

with very high probability 1 − γ.

Lemma 5.11. Let the field F, degree d and µ, γ, ε ∈ (0, 1) satisfy q ≥ 800d · max {1/(µγ)16 , 1/µγ 2 ε16 }.
There is a integer C and τ ∈ (0, 1) such Given a function f and ε-well-behaved degree-d lines oracle,
let S be a subset of the 2ε-good points in Fm of density µ. Then, the distribution (x, ℓ) obtained by
picking a random point in S and a random line ℓ in Fm satisfies
h i
x
Pr ∃ degree-d polynomial Pℓ such that δ(fcorr |ℓ , P ) ≤ γ ≥ 1 − 2γ.
x∈S,ℓ ε8 /2

We now complete the proof of Theorem 5.7 assuming this lemma.

Proof of Theorem 5.7. We are given a points table f and a degree-d lines oracle ℓ 7→ Pℓ for m
dimensions such that Prx,ℓ∋x [f (x) = Pℓ (x)] ≥ 5ε. We first modify the lines oracle by setting Pℓ
to ⊥ if Prx∈ℓ [f (x) = Pℓ (x)] < ε. This ensures, that the lines oracle is ε-well-behaved wrt f . This
reduces the acceptance probability of the LDT by at most ε and the modified lines oracle satisfies

Pr [f (x) = Pℓ (x)] ≥ 4ε. (5.12)

x,ℓ∋x

We will be setting µ := ε and γ := ε2/12. There exists a suitable large constant C, such that
q≥ C·d/ε48 implies that the field-size is large enough for Lemma 5.14 for this choice of ε, µ, γ.
(ℓ) (ℓ)
For any line ℓ, let P1 , . . . , Pr be all the univariate degree-d polynomials P that satisfy
Prz∈ℓ [P (z) = f (z)] ≥ ε8/2. Let nonuniquedε8/2 (f |ℓ ) be the set of all points x on the line ℓ such
(ℓ) (ℓ) (ℓ) (ℓ)
that there exist two distinct polynomials Pk 6= Pk′ such that Pk (x) = Pk′ (x). By Theorem 5.1-

30
2, we have that the number of points in nonuniquedε8/2 (f |ℓ ) is at most 8d/ε16 .

The LDT hypothesis (5.12) implies that

/ nonuniquedε8/2 (fℓ )] ≥ 4ε − 8d/q·(ε8 )2 ≥ 3ε,

Pr [f (x) = Pℓ (x) and x ∈
ℓ,x∈ℓ

provided q ≥ 8d/ε·ε16 . Or equivalently,

 h i
d
E Pr f (x) = Pℓ (x) and x ∈
/ nonuniqueε8/2 (f |ℓ ) ≥ 3ε.
x ℓ∋x

Define S ⊆ Fm to be the set of points as follows:

 h i 
m d
S := x ∈ F : Pr f (x) = Pℓ (x) and x ∈
/ nonuniqueε8/2 (f |ℓ ) ≥ 2ε .
ℓ∋x

We thus have Prx∈Fm [x ∈ S] ≥ ε. Also observe that any x ∈ S is 2ε-good. We now apply
x
Lemma 5.11 on the set S with µ := ε to obtain that fcorr passes the LDT with probability at
ε8/2
least 3γ for a random x. Fix any such x ∈ S. We have δfcorr
x ≤ 3γ where δf is as defined in
ε8/2

(5.6). Applying Theorem 1.2 (m-variate LDT in the low-error regime), we obtain that there exists
a m-variate degree-d polynomial Qx such that δ(Qx , fcorr
x ) ≤ 12γ.
ε8 /2
We will now argue that f and x
fcorr agree on at least 2ε2 fraction of points since x ∈ S.
ε8 /2
x
Recall the definition of S and the definition of the x-corrected function fcorr . For every line ℓ
ε8/2

/ nonuniquedε8/2 , we have that fcorr

through x such that f (x) = Pℓ (x) and x ∈ x (z) = Pℓ (z) for
ε8/2

every z ∈ ℓ such that Pℓ (z) = f (z). This is because (1) Pℓ agrees with f on at least ε ≥ ε8/2-

fraction of the points (since the lines oracle is ε-well-behaved), (2) Pℓ (x) = f (x) and (3) Pℓ is the
x
unique polynomial satisfying (1) and (2). Hence, for any such line ℓ, we have Prz∈ℓ [fcorr (z) =
ε8/2

f (z)] ≥ ε. Furthermore, there are at least 2ε-fraction of such lines through x since x ∈ S. Hence,
x
Prz [fcorr (z) = f (z)] ≥ 2ε · ε = 2ε2 .
ε8/2

Combining this, with δ(Qx , fcorr

x ) ≤ 12γ, we have that Prz [Qx (z) = f (z)] ≥ 2ε2 − 12γ ≥ ε2
ε8 /2
provided γ ≤ ε2/12. This completes the proof of the theorem.

5.4.1 Corrected function passes LDT with high probability (proof of Lemma 5.11)

Our plan is to show that if the subset S of 2ε-globally good points is of density at least µ, then
a random point x in S satisfies that the corresponding x-corrected function passes the low-degree
test with probability 1 − O(γ) (for arbitrarily small γ > 0 provided d/q is small enough). To do so
we consider a random point x ∈ S and a random line ℓ and consider the plane π containing x and
ℓ and prove that with probability 1 − γ the following hold: There is a bivariate polynomial Qxπ on
the plane π such that (1) Qxπ has agreement at least ε8 with f |π , (2) Qxπ (x) = f (x), and (3) For

31
1 − γ fraction of the points y ∈ π, if we let ℓ′ be the line through x and y, then Qxπ |ℓ′ is the unique
polynomial with ε8 /2 agreement with f |ℓ′ .
Note that the existence of even one point x with properties (1)-(3) above suffice, but the proof
essentially forces us to prove that most points in S satisfy properties (1)-(3).
We start with a preliminary statement. Note that the distribution on the triples (x, ℓ, π) of
interest to us is the following: D1 is the distribution obtained by picking x to be a uniform point
in S, ℓ a uniformly random line in Fm and π be a uniform plane containing ℓ and x (w.h.p. π is
unique given ℓ and x, but if x ∈ ℓ then π is a uniform plane containing ℓ) and output (x, ℓ, π). Now
consider the related distribution D2 obtained by sampling a plane π uniformly in Fm , then sampling
a line ℓ uniformly in π and a point x ∈ S on π and outputting (x, ℓ, π). If there is no point of S
in π (i.e., S ∩ π = ∅), return ⊥. The following claim establishes that these two distributions are
O(1/(εq))-close in statistical distance.

Claim 5.13. Suppose |F| ≥ 3/γ · √µ where µ = Prx [x ∈ S]. Then kD1 − D2 kTV ≤ γ

f1 and D
Proof. We construct two related distributions D f2 as follows. D
f1 is the distribution obtained
by picking x to be a uniform point in S, ℓ a uniformly random line in Fm not containing x in Fm
and π be a uniform plane containing ℓ and x (note π is unique given ℓ and x) and output (x, ℓ, π).
f2 is obtained by sampling a plane π uniformly in Fm , a point x ∈ S on π,
The distribution D
a line ℓ uniformly in π not containing x and then outputting (x, ℓ, π). If there is no point of
S in π (i.e., S ∩ π = ∅), return ⊥. The only difference between D1 and D f1 is that the line ℓ
f1 while it may do so with probability with probability 1/qm−1
definitely does not pass through x in D
in D1 . Thus, kD1 − D f1 kTV ≤ 1/qm−1 . Similarly, kD2 − D
f2 kTV ≤ 1/q. Hence, kD1 − D2 kTV ≤
kDf1 − Df2 kTV + 1/q + 1/qm−1 ≤ 2/q.
f1 . It can alternatively be sampled as follows. Pick a random point
Consider the distribution D
x ∈ S, a random plane π containing x, a line ℓ uniformly in π not containing x and then outputting
f1 and D
the triple (x, ℓ, π). Since the generative processes for picking ℓ in D f2 are identical given the
f1 and D
pair (x, π) which are picked differently, the distance between the distributions D f2 is exactly
the distance between the marginals of D f1 and Df2 on the the (x, π) coordinates. By Lemma 5.4,
this distance is at most λ/µ(S) where λ is the second eigen-value of the points-plane incidence graph
G = G(Fm , P (m) ) which is at most 1/q (Theorem 5.5-2). Hence, kD f2 kTV ≤ 1/q√µ.
f1 − D
Hence, kD1 − D2 kTV ≤ 2/q + q/q√µ ≤ γ provided q ≥ 3/γ · √µ.

Given this claim, it suffices to prove the following lemma to prove Lemma 5.11.

Lemma 5.14. Let the field F, degree d and µ, γ, ε ∈ (0, 1) satisfy q ≥ 800d · max {1/(µγ)16 , 1/µγ 2 ε16 }.
Given a function f and ε-well-behaved degree-d lines oracle, let S be a subset of the 2ε-good points
in Fm of density µ. Then, the distribution D2 on triples (π, x, ℓ) obtained by picking a random plane

32
π in Fm , a random x ∈ S in the plane π and a random line ℓ in the plane πsatisfies
h i
Pr ∃ degree-d polynomial Pℓx such that δ(fcorr
x
|ℓ , Pℓx ) ≤ γ ≥ 1 − γ .
(π,x,ℓ)∼D2 ε8 /2

Proof of Lemma 5.11. Follows from Claim 5.13 and Lemma 5.14

Proof. To prove Lemma 5.14, we list some bad events E1–E4 such that if none of these occur then
x
fcorr passes the low-degree test on (ℓ, y) with probability 1 − 2γ over the choice of y ∈ ℓ. We
ε8 /2
then argue that each of these events happens with probability at most γ/4.

Event E1(π): Prz∈π [z ∈ S] < µ/2.

Event E2(π): Prz∈π [z ∈ S but z is not ε-locally-good wrt. plane π] ≥ µ·γ/16.

Event E3(π, x): ¬(E1 ∨ E2) and “x is not ε8 -explained on π”.

Now if the plane π and point x are such that E1, E2 and E3 do not occur, then there exists
a bivariate degree-d polynomial Qxπ on the plane π such that Prz∈π [Qxπ (z) = f (x)] ≥ ε8 and
f (x) = Qxπ (x).
x
Event E4(π, x, ℓ): ¬(E1 ∨ E2 ∨ E3) and Pry∈ℓ [fcorr (y) 6= Qxπ (y)] > γ.
ε8 /2

Claims 5.15, 5.16, 5.18 and 5.19 imply that each of these events occurs with probability at most
γ/4. Furthermore, clearly if none of the 4 events occur, then the polynomial Qxπ disagrees with the
x
corrected function fcorr on at most γ-fraction of the points on the line ℓ. This completes the
ε8 /2
proof of the lemma assuming Claims 5.15, 5.16, 5.18 and 5.19.

The rest of this section is devoted towards bounding the probability of the events E1–E4.

Claim 5.15 (Bounding E1). If µ = Prx [x ∈ S] and |F| ≥ 4/√µ·γ , then

Pr[E1] = Pr[|S ∩ π| < µ/2 · q 2 ] ≤ γ/4.

π π

Proof. Consider the bipartite points-planes incidence graph G(Fm , P (m) ) in Fm which has second
eigen-value at most 1/q (by Theorem 5.5-2). Consider the functions g : Fm → R and h : P (m) → R
defined as follows: g(z) := 1[z ∈ S] and h(π) := 1[|S ∩ π| < µ/2 · q 2 ]. These functions satisfy µg = µ,
µh = Pr[E1] =: α, σg2 ≤ µ and σh2 ≤ α. Applying Lemma 5.3 to the graph G with functions g and
h as defined above, we have

√
µ · α − µ/2 · α ≤ λ · µ · α.

Equivalently, α ≤ 4λ2/µ, which in turn is at most γ/4 (since λ ≈ 1/|F| and |F| ≥ 4/√µ·γ ).

33
Claim 5.16 (Bounding E2). If |F| + 1 ≥ 100/ε·γ 2 and µ = Prx [x ∈ S], then

Pr[E2] = Pr[π has at least µ·γ/16 · q 2 points which are in S but not ε-locally-good ] ≤ γ/4.
π π

Proof. We begin by showing the following: If for any 2ε-good point x (in particular, if x ∈ S) and
|F| + 1 ≥ 100/ε·γ 2 , then

Pr [x is not ε-good with respect to plane π] ≤ γ 2/64. (5.17)

π∈Px

(m)
Let x be an 2ε-good point. Let Lx ⊆ Lx be the set of lines ℓ containing x that satisfy
Pℓ (x) = f (x). We have Prℓ∋x [ℓ ∈ Lx ] ≥ 2ε. Consider the bipartite lines-plane incidence graph
(m) (m)
G(Lx , Px ) in Fm between lines and planes containing x. This graph has second eigen-value at
√ (m) (m)
most 1/ q+1 (by Theorem 5.5-4). Consider the functions g : Lx → R and h : Px → R defined
as follows: g(ℓ) := 1[f (x) = Pℓ (x)] = 1[ℓ ∈ Lx ] and h(π) := 1[Prℓ∈π [ℓ ∈ Lx ] < ε]. These functions
satisfy µh =: α, σg2 ≤ µg and σh2 ≤ α. Applying Lemma 5.3 to the graph G with functions g and h
as defined above, we have

√
µg · α − ε · α ≤ λ · µg · α.

Equivalently, α ≤ λ2 ·µg/(µg −ε)2 which is at most λ2/ε (since µg ≥ 2ε), which in turn is at most γ 2/64
√
(since λ ≈ 1/ |F|+1 and |F| + 1 ≥ 64/ε·γ 2 ), completing the proof of (5.17).
We now return to the proof of Claim 5.16. Consider the bipartite point-plane inclusion graph
G = G(Fm , P (m) ). Let A := S and µ := Prz [z ∈ A]. Let B ⊂ P (m) be the set of planes π which
have at least µ·γ/16 · q 2 points which are in S but not ε-locally-good. We mark an edge (x, π) if
x ∈ A (i.e., x is 2ε-globally-good) but x is not ε-good with respect to the plane π. The fraction
of marked edges in G by Eq. (5.17) is at most µ · γ 2/64. On the other hand, the fraction of marked
edges is at least µ(B) · µ·γ/16. Hence, µ(B) · µ·γ/16 ≤ µ · γ 2/64 or equivalently, µ(B) ≤ γ/4.

Claim 5.18 (Bounding E3). If the parameter ε0 in (5.10) satisfies ε0 ≤ µ·γ/16 where µ = Prz [z ∈ S],
then Prπ,x [E3] ≤ γ/4.

Proof. To begin with fix a plane π such that neither E1 nor E2 occurs. Hence, π has at least
µ/2 · q 2 points in S of which at most µ·γ/16 · q 2 are not ε-locally-good. Hence, there are at least
µ·(1−γ/8)/2· q 2 ≥ µ/4 · q 2 points which are ε-locally-good in π.
Now, by the bivariate LDT (5.10), we have that the probability that a random x ∈ Fm is ε-
locally-good but not ε8 -explained (both with respect to π) is at most ε0 . Hence, the probability
that a random point in S ∩ π is not ε8 -explained is at most

µ·γ/16+ ε0 µ·γ/16 + ε0
≤ = γ/8 + 2ε0/µ ≤ γ/4,
µ (S ∩ π) µ/2

34
where the last inequality follows if ε0 ≤ µ·γ/16. We now bound E3 as follows:

Pr[E3] ≤ E[Pr[E3|¬(E1 ∨ E2)]] = E[ Pr [x is not ε8 -explained |¬(E1 ∨ E2)]] ≤ γ/4.

π,x π x π x∈π

Claim 5.19 (Bounding E4). If |F| ≥ 800d/µ·γ 2 ·ε16 , then Prπ,x,ℓ [E4] ≤ γ/4.x

Proof. The argument for bounding Pr[E4] will be far more involved and subtler than the previous
cases.
Let us for the rest of the argument fix a plane π. Furthermore, let us assume that this plane π
and a random x ∈ S chosen on it are such that none of E1, E2 and E3 occur. This implies that π
has at least µ/2-fraction of points which are in S and there exists a bivariate degree-d polynomial
Qxπ such that (1) Prz∈π [Qxπ (z) = f (z)] ≥ ε8 and (2) Qxπ (x) = f (x). We now need to bound the
probability that when we choose a random line ℓ in the plane π, at least γ-fraction of the points y
on ℓ satisfy Qxπ (y) 6= fcorr
x x
(y). To this end, let us recall the definition of fcorr : Fm → F ∪ {⊥}.
ε8 /2 ε8 /2
For any point y ∈ Fm \ {x}, let P be the unique degree-d univariate polynomial P on the line
ℓ′ = ℓx,y through x and y such that (1) Prz∈ℓ′ [P (z) = f (z)] ≥ ε8 /2 and (2) P (x) = f (x). If there is
no such polynomial P or there is more than one such polynomial, then we set P to ⊥. Finally, we
x
set fcorr x
(y) := P (y) and fcorr (x) := f (x). We will now argue that the probability that this
ε8 /2 ε8 /2
polynomial P is Qxπ |ℓ′ for at least 1 − γ fraction of the points y in ℓ (this probability will be over
the random choice of an ε-good point x in π and a random line ℓ in π).
Let Q1 , . . . , Qt be the list of all degree-d bivariate polynomials on π such that Pry∈π [Qj (y) =
p
f (y)] ≥ ε8 . By Johnson bound Theorem 5.1-1, we have t ≤ 2/ε8 provided ε8 ≥ 2 d/q. Note that
Qxπ is one such polynomial. For j ∈ [t], let Sj := {x ∈ π : Qj (x) = f (x)} be the set of agreement
points between f and the polynomial Qj (by definition, µ(Sj ) ≥ ε8 ).
We list below some bad events B1, B2 and B3, which if they do not occur would imply that for
a random x ∈ S, a random line ℓ and a point y ∈ ℓ, the polynomial P is the restricted polyomial
Qxπ |ℓ′ (where ℓ′ = ℓx,y ). These bad events would be described over the randomness of the choice of
the point x ∈ S and the random line ℓ′ ∈ Lxπ (note ℓ′ = ℓx,y is a random line through the point x
in the plane π)

Event B1(ℓ′ ): There exists j ∈ [t] such that |Sj ∩ ℓ′ | < ε8/2 · q.

Event B2(ℓ′ ): |S ∩ ℓ′ | < µ/4 · q.

Let P1 , . . . , Pr be the list of all univariate degree-d polynomials on the line ℓ′ such that
Prz∈ℓ′ [Pk (z) = f (z)] ≥ Note if B1 does not occur, then Qxπ |ℓ′ is one such polyno-
ε8/2.
p
mial. By Johnson bound Theorem 5.1, we have r ≤ 4/ε8 provided ε8/2 ≥ 2 d/q . We say x
is a non-unique point on ℓ′ if there exist two distinct polynomials k 6= k ′ ∈ [r] such that
Pk (x) = Pk′ (x).

Event B3(ℓ′ , x): ¬B2 and x is a non-unique point on ℓ′ .

35
Applying Claim 5.20 with δ := γ 2/8, we have Pr(ℓ′ ,x)∼E [B1∨B2∨B3] ≤ γ 2/8 provided q ≥ 800d/µ·γ 2 ·ε16 .
However, the distribution E over the pairs (x, ℓ′ ) is slightly different from what we need. Our
required distribution of (x, ℓ′ ) is as follows: given a plane π, pick a random point x ∈ S on it and a
random line ℓ′ in π passing through x. The distribution E , on the other hand, is as follows: given a
plane π, pick a random line ℓ′ and a random x ∈ S on ℓ′ (if one exists). By Lemma 5.4, these two
√
distributions are λ/ µ(S∩π)-close in total variation distance where λ ≈ 1/√q is the second eigen-value
of the points-line incidence graph in the plane π (Lemma 5.4-1) and µ (S ∩ π) ≥ µ/2 (since event E1
p
does not hold). Hence, this distance is at most 2/q·µ which is at most γ 2/8 provided q ≥ 200/µ·γ 4 .
Hence, Prx,ℓ′ [B1 ∨ B2 ∨ B3] ≤ γ 2/4.
Now, recall that π and x are such that none of the events E1–E3 occur. Hence, π has at least
µ/2-fraction of 2ε-good points and there exists a bivariate degree-d polynomial Qxπ such that (1)
Prz∈π [Qxπ (z) = f (z)] ≥ ε8 and (2) Qxπ (x) = f (x). Now, suppose furthermore that ℓ and y ∈ ℓ (and
the corresponding ℓ′ = ℓx,y ) are such that B1–B3 do not occur. It follows from these assumptions
that Qxπ is one of the polynomials Qj and Qxπ |ℓ′ one of the polynomials Pk . The uniqueness condition
of ¬B3 implies that Qxπ |ℓ′ is the only degree-d polynomial that has agreement at least ε8/2 with f
on ℓ and Qxπ (x) = f (x). Hence, x
fcorr (y) = Qxπ (y). We have thus shown that
ε8
/2

x
Pr [fcorr (y) 6= Qxπ (y)] ≤ γ 2/4.
x,ℓ,y∈ℓ ε8/2

x
We are however interested in the fraction of points in ℓ such that fcorr (y) 6= Qxπ (y). A Markov
ε8/2
argument shows that
 
x
Pr {y ∈ ℓ : fcorr (y) 6= Qxπ (y)} > γ · q ≤ γ 2 ·q/4·γ·q = γ/4.
x,ℓ ε8/2

Averaging over π, yields the claim.

Claim 5.20. Let ρ, δ, ε ∈ (0, 1), field F of size q and degree parameter d satisfy q ≥ 100d/µ·δ·ε16 . For
any plane π and a set S of 2ε-good points in the plane of density at least µ/2, consider the distribution
E on pairs (ℓ′ , x) chosen as follows: ℓ′ is a random line in the plane π and x is a random point on
S ∩ ℓ (if not such point exists, then the distribution returns ⊥). Then, Pr(ℓ,x′ )∼E [B1 ∨ B2 ∨ B3] ≤ δ.

Proof. We bound the probability of each of the events B1, B2 and B3 by δ/3 as follows:

Event B1(ℓ′ ): There exists j ∈ [t] such that |Sj ∩ ℓ′ | < ε8/2 · q.
For any fixed j ∈ [t], since µ(Sj ) ≥ ε8 and the set of points in a random line ℓ′ are pairwise
independent, we have Pr[|Sj ∩ ℓ′ | < ε8/2 · q] ≤ q·ε8/(q·ε8 /2)2 = 4/qε8 . Hence, Pr[B1] ≤ t · 4/qε8 ≤
8/q·ε16 (as t ≤ 2/ε8 ). We, hence, have Pr[B1] ≤ δ/3 provided q ≥ 24/δ·ε16

Event B2(ℓ′ ): |S ∩ ℓ′ | < µ/4 · q.

36
 Since µ (S ∩ π) ≥ µ/2 and the set of points in a random line ℓ′ are pairwise independent,
we have Pr[B2] = Pr[|S ∩ ℓ′ | < µ/4 · q] ≤ q·µ/2/(q·µ/4)2 = 8/qµ, which is at most δ/3 provided
q ≥ 24/µ·δ .

Event B3(ℓ′ , x): ¬B2 and x is a non-unique point on ℓ′ .

Since event B2 does not happen, we know that that at least µ/4-fraction of points in ℓ′ are
2ε-good. The fraction of points x in ℓ such that there exist two distinct polynomials Pk Pk′


which coincide on x (i.e., Pk (x) = Pk′ (x)) is at most r2 · d/q ≤ 8d/q·(ε8 )2 since r ≤ 4/ε8 . Hence,
the probability of event B3 that a random point in ℓ′ ∩ S happens to be non-unique is at most
4·8d/µ·q·(ε8 )2 . Hence, Pr[B3] ≤ δ/3 provided q ≥ 100d/µ·δ·ε16 .

Acknowledgements
Some of the discussions of the first two authors leading up to this work happened while they were
visiting the Homi Bhabha Center for Science Education (HBCSE), Mumbai. We are thankful to
Prof. Arnab Bhattacharya and rest of the HBCSE staff for their warm and generous hospitality
and for providing an inviting and conducive atmosphere for these discussions.

References
[AKKLR05] Noga Alon, Tali Kaufman, Michael Krivelevich, Simon Litsyn, and Dana Ron.
Testing Reed-Muller codes. IEEE Trans. Inform. Theory, 51(11):4032–4039, 2005. (Preliminary
version in 7th RANDOM, 2003).
(Referenced on page 4.)

[ALMSS98] Sanjeev Arora, Carsten Lund, Rajeev Motwani, Madhu Sudan, and Mario
Szegedy. Proof verification and the hardness of approximation problems. J. ACM, 45(3):501–
555, May 1998. (Preliminary version in 33rd FOCS, 1992). eccc:1998/TR98-008.
(Referenced on page 3.)

[And20] Robert Andrews. Algebraic hardness versus randomness in low characteristic. In Shubhangi
Saraf, ed., Proc. 35th Comput. Complexity Conf., volume 169 of LIPIcs, pages 37:1–37:32.
Schloss Dagstuhl, 2020. arXiv:2005.10885, eccc:2020/TR20-081.
(Referenced on page 9.)

[AS98] Sanjeev Arora and Shmuel Safra. Probabilistic checking of proofs: A new characterization
of NP . J. ACM, 45(1):70–122, January 1998. (Preliminary version in 33rd FOCS, 1992).
(Referenced on page 3.)

[AS03] Sanjeev Arora and Madhu Sudan. Improved low-degree testing and its applica-
tions. Combinatorica, 23(3):365–426, 2003. (Preliminary version in 29th STOC, 1997).
eccc:1997/TR97-003.
(Referenced on pages 3, 4, 5, 9, 10, 11, 12, 19, 24, 26, 41, and 48.)

37
[BCIKS20] Eli Ben-Sasson, Dan Carmon, Yuval Ishai, Swastik Kopparty, and Shubhangi Saraf.
Proximity gaps for Reed-Solomon codes. In Sandy Irani, ed., Proc. 61st IEEE Symp. on
Foundations of Comp. Science (FOCS), pages 900–909. 2020. eccc:2020/TR20-083.
(Referenced on page 6.)

[BDN17] Amey Bhangale, Irit Dinur, and Inbal Livni Navon. Cube vs. Cube low degree test.
In Christos Papadimitriou, ed., Proc. 8th Innovations in Theor. Comput. Sci. (ITCS),
volume 67 of LIPIcs. Schloss Dagstuhl, 2017. arXiv:1612.07491, eccc:2016/TR16-205.
(Referenced on pages 5, 10, 11, 24, 25, and 26.)

[BGHSV05] Eli Ben-Sasson, Oded Goldreich, Prahladh Harsha, Madhu Sudan, and Salil
Vadhan. Short PCPs verifiable in polylogarithmic time. In Luca Trevisan, ed., Proc.
20th IEEE Conf. on Comput. Complexity, pages 120–134. 2005. Full version available at
http://www.tcs.tifr.res.in/~prahladh/papers/BGHSV2/BGHSV2005.pdf.
(Referenced on page 4.)

[BGHSV06] ———. Robust PCPs of proximity, shorter PCPs and applications to coding. SIAM J. Comput.,
36(4):889–974, 2006. (Preliminary version in 36th STOC, 2004). eccc:2004/TR04-021.
(Referenced on page 4.)

[BHKS23] Siddharth Bhandari, Prahladh Harsha, Mrinal Kumar, and Madhu Sudan. Decoding
multivariate multiplicity codes over product sets. IEEE Trans. Inform. Theory, 2023. (Early
Access, Preliminary version in 53rd STOC, 2021). arXiv:2012.01530, eccc:2020/TR20-179.
(Referenced on page 9.)

[BSVW03] Eli Ben-Sasson, Madhu Sudan, Salil Vadhan, and Avi Wigderson. Randomness-effi-
cient low degree tests and short PCPs via epsilon-biased sets. In Lawrence L. Larmore and
Michel X. Goemans, eds., Proc. 35th ACM Symp. on Theory of Computing (STOC), pages
612–621. 2003.
(Referenced on page 4.)

[Bür00] Peter Bürgisser. Completeness and Reduction in Algebraic Complexity Theory, volume 7 of
Algorithms and computation in mathematics. Springer, 2000.
(Referenced on page 15.)

[Cha16] Siu On Chan. Approximation resistance from pairwise-independent subgroups. J. ACM,

63(3):27:1–27:32, 2016. (Preliminary version in 45th STOC, 2013). eccc:2012/TR12-110.
(Referenced on page 5.)

[DL78] Richard A. DeMillo and Richard J. Lipton. A probabilistic remark on algebraic program
testing. Inform. Process. Lett., 7(4):193–195, 1978.
(Referenced on page 13.)

[FS95] Katalin Friedl and Madhu Sudan. Some improvements to total degree tests. In Proc.
3rd Israel Symp. on Theoretical and Computing Systems, pages 190–198. 1995. (See arXiv for
corrected version). arXiv:1307.3975.
(Referenced on pages 3, 5, 6, 7, 10, 11, 24, and 25.)

38
[GKSS22] Zeyu Guo, Mrinal Kumar, Ramprasad Saptharishi, and Noam Solomon. Derandom-
ization from algebraic hardness. SIAM J. Comput., 51(2):315–335, 2022. (Preliminary version
in 60th FOCS, 2019). arXiv:1905.00091, eccc:2019/TR19-065.
(Referenced on page 9.)

[GW13] Venkatesan Guruswami and Carol Wang. Linear-algebraic list decoding for variants of
Reed-Solomon codes. IEEE Trans. Inform. Theory, 59(6):3257–3268, 2013. (Preliminary version
in 26th IEEE Conference on Computational Complexity, 2011 and 15th RANDOM, 2011).
eccc:2012/TR12-073.
(Referenced on page 9.)

[Kal89] Erich L. Kaltofen. Factorization of polynomials given by straight-line programs. In Silvio

Micali, ed., Randomness and Computation, volume 5 of Advances in Computing Research,
pages 375–412. JAI Press, Greenwich, Connecticut, 1989.
(Referenced on page 9.)

[Kop14] Swastik Kopparty. Some remarks on multiplicity codes. In Alexander Barg and Oleg R.
Musin, eds., Discrete Geometry and Algebraic Combinatorics, volume 625 of Contemporary
Mathematics, pages 155–176. AMS, 2014. arXiv:1505.07547.
(Referenced on page 9.)

[Kop15] ———. List-decoding multiplicity codes. Theory of Computing, 11:149–182, 2015.

eccc:2012/TR12-044.
(Referenced on page 9.)

[KSS15] Swastik Kopparty, Shubhangi Saraf, and Amir Shpilka. Equivalence of polynomial
identity testing and polynomial factorization. Comput. Complexity, 24(2):295–331, 2015. (Pre-
liminary version in 29th CCC, 2014). eccc:2014/TR14-001.
(Referenced on page 9.)

[LN96] Rudolf Lidl and Harald Niederreiter. Finite Fields, volume 2 of Encyclopedia of Math-
ematics and its Applications. Cambridge University Press, 2nd edition, 1996.
(Referenced on page 40.)

[MR08] Dana Moshkovitz and Ran Raz. Sub-constant error low degree test of almost-linear
size. SIAM J. Comput., 38(1):140–180, 2008. (Preliminary version in 38th STOC, 2006).
eccc:2005/TR05-086.
(Referenced on page 4.)

[MR10a] ———. Sub-constant error probabilistically checkable proof of almost-linear size. Comput.
Complexity, 19(3):367–422, 2010. eccc:2007/TR07-026.
(Referenced on page 4.)

[MR10b] ———. Two-query PCP with subconstant error . J. ACM, 57(5), 2010. (Preliminary version in
49th FOCS, 2008). eccc:2008/TR08-071.
(Referenced on page 4.)

[MZ23] Dor Minzer and Kai Zheng. Approaching the soundness barrier: A near optimal anal-
ysis of the cube versus cube test. In Nikhil Bansal and Viswanath Nagarajan, eds.,

39
 Proc. 34th Annual ACM-SIAM Symp. on Discrete Algorithms (SODA), pages 2761–2776. 2023.
arXiv:2211.09341.
(Referenced on page 5.)

[Ore22] Øystein Ore. Über höhere kongruenzen (German) [About higher congruences]. Norsk Mat.
Forenings Skrifter, 1(7):15, 1922. (see [LN96, Theorem 6.13]).
(Referenced on page 13.)

[PS94] Alexander Polishchuk and Daniel A. Spielman. Nearly-linear size holographic proofs.
In Frank Thomson Leighton and Michael T. Goodrich, eds., Proc. 26th ACM Symp.
on Theory of Computing (STOC), pages 194–203. 1994.
(Referenced on pages 3, 4, 6, and 9.)

[RS96] Ronitt Rubinfeld and Madhu Sudan. Robust characterizations of polynomials with appli-
cations to program testing. SIAM J. Comput., 25(2):252–271, April 1996. (Preliminary version
in 23rd STOC, 1991 and 3rd SODA, 1992).
(Referenced on pages 3, 5, and 10.)

[RS97] Ran Raz and Shmuel Safra. A sub-constant error-probability low-degree test, and a sub–
constant error-probability PCP characterization of NP . In Frank Thomson Leighton and
Peter W. Shor, eds., Proc. 29th ACM Symp. on Theory of Computing (STOC), pages 475–
484. 1997.
(Referenced on pages 3, 4, 5, 10, and 11.)

[Sch80] Jacob T. Schwartz. Fast probabilistic algorithms for verification of polynomial identities. J.
ACM, 27(4):701–717, October 1980.
(Referenced on page 13.)

[Ska13] Matthew Skala. Hypergeometric tail inequalities: ending the insanity, 2013. (manuscript).
arXiv:1311.5939.
(Referenced on page 42.)

[STV01] Madhu Sudan, Luca Trevisan, and Salil P. Vadhan. Pseudorandom generators without
the XOR lemma. J. Comput. Syst. Sci., 62(2):236–266, 2001. (Preliminary version in 31st
STOC, 1999).
(Referenced on pages 5, 11, and 24.)

[Zip79] Richard Zippel. Probabilistic algorithms for sparse polynomials. In Edward W. Ng, ed.,
Proc. International Symp. of Symbolic and Algebraic Computation (EUROSAM), volume 72 of
LNCS, pages 216–226. Springer, 1979.
(Referenced on page 13.)

A An Arora-Sudan style interpolation

In this section, we present the full proof of Theorem 4.3. We repeat the statement below for
convenience.
git info: 542509c , (2023-11-21 22:56:27 +0530) (What’s the point of this line?)

40
Theorem 4.3 (Interpolation for LDT). There exist constant c1 , c2 ∈ N such that for every ε ∈ (0, 1],
any d ∈ N and finite field F of size q and characteristic p with q > c1 · d/εc2 , the following is true.
Let f : F2 → F be an oracle that passes LDTd with probability ε. Then, there is a non-zero
polynomial A(x, y, z) and a subset S ⊆ F2 such that

• deg1,1,d (A) = O(d/ε2 ) and |S| = Ω(ε2 q 2 ),

P 3 2
• x∈S εx = Ω(ε · q ),

• for every (a, b) ∈ S, we have ε(a,b) = Prℓ∋(a,b) [Pℓ (a, b) = f (a, b)] ≥ ε/2,
• for every (a, b) ∈ S, we have A(a, b, f (a, b)) = 0.
• ∂¯ z (A) and Discz (A) are not identically zero.

The proof is mostly along the lines of Arora and Sudan [AS03] but a tighter analysis and some
care for fields of small characteristic. We present it in its entirety for completeness and to make the
changes clearer to follow.

A.1 Finding a structured subset of points

We start with the following claim.

Claim A.1 (Finding good directions). Let q > 8/ε and suppose f : F2 → F passes LDTd with
probability ε. Then, there are two different directions ℓ1 and ℓ2 , and a set H ⊆ F2 such that

• |H| = Ω(ε2 · q 2 ),
• for all x ∈ H we have Prℓ∋x [Pℓ (x) = f (x)] ≥ ε/2 and

Pℓ(x) (x) = Pℓ(x) (x) = f (x)

1 2

(x)
where ℓi is the line through x parallel to ℓi .
P
Proof. For each x ∈ F2 , let εx = Prℓ∋x [Pℓ (x) = f (x)] and we have x εx = ε · q 2 . Let H ′ =
{x : εx ≥ ε/2}. For a direction ℓ and a point x ∈ F2 , let I(x, ℓ) be indicator random variable
defined as

I(x, ℓ) = 1 [Pℓ(x) (x) = f (x) and εx ≥ ε/2]

where ℓ(x) is the line through x parallel to ℓ. We then have Eℓ [I(x, ℓ)] = εx if x ∈ H ′ and 0
otherwise. Thus,
X
E E[I(x, ℓ)] = 1/q2 · εx ≥ ε/2 (by Lemma 2.1).
x ℓ
x∈H ′

41
Therefore, we have

ε/2 ≤ E E[I(x, ℓ)]

x ℓ
 2
ε /4 ≤ E E[I(x, ℓ)]
2
≤ E(E[I(x, ℓ)])2
x ℓ x ℓ
 
= E E [I(x, ℓ1 ) · I(x, ℓ2 )]
x ℓ1 ,ℓ2
 
1 1
=E E [I(x, ℓ1 ) · I(x, ℓ2 )] + 2 (∵ Pr[ℓ = ℓ′ ] = )
x ℓ1 6=ℓ2 q +q q2 +q
=⇒ ε2/8 ≤ E E [I(x, ℓ1 ) · I(x, ℓ2 )] (∵ q > 8/ε)
x ℓ1 6=ℓ2

Therefore, there exist two different directions ℓ1 and ℓ2 such that Ex [I(x, ℓ1 )I(x, ℓ2 )] ≥ ε2/8. Fixing
such directions ℓ1 and ℓ2 , defining H = {x : I(x, ℓ1 ) = I(x, ℓ2 ) = 1} satisfies the requirements.

Without of loss of generality, we assume that ℓ1 is the x-axis and ℓ2 is the y-axis. Let γ be
chosen such that H = 2γ · q 2 = Ω(ε2 q 2 ).

Lemma A.2 (Structured subsets within H). Let H be the set specified above with respect to the x
and y directions for an oracle f passing LDTd with probability at least ε. Then, for any r satisfying
2 log q
γ2
≤ r ≤ γ · q, there are subsets S1 , S2 ⊆ F such that

1. |S1 | = r,
2. |S2 | ≥ γ · q,
3. |(F × S2 ) ∩ H| ≥ |H|/2 = γ · q 2 ,
4. For all b ∈ S2 , we have |{(a, b) : a ∈ F} ∩ H| ≥ γ · q,
5. For all b ∈ S2 , we have |{(a, b) : a ∈ S1 } ∩ H| ≥ γ/2 · |S1 |.

Proof. Let ℓy=b = {(a, b) : a ∈ F} and let ℓx=a = {(a, b) : b ∈ F}. Since these lines evenly cover
the space, we have that Eb [|ℓy=b ∩ H|] = 2γ ·q. Let S2 = {b ∈ F : |ℓy=b ∩ H| ≥ γq}. By Lemma 2.1,
we have that |S2 | ≥ γ · q and |(F × S2 ) ∩ H| ≥ |H|/2 = γ · q 2 .
Let S1 be a set of r distinct elements of F chosen uniformly at random. Then, for any b ∈ S2

E [|{(a, b) : a ∈ S1 , (a, b) ∈ H}|] = r/q · |ℓy=b ∩ H| ≥ γ · r.

S1

From standard tail bounds for hypergeometric distributions (cf. [Ska13]), we have

For all b ∈ S2 , Pr [|{(a, b) : a ∈ S1 , (a, b) ∈ H}| < γr/2] ≤ exp(−γ 2 r/2)

S1
=⇒ Pr [∃b ∈ S2 : |{(a, b) : a ∈ S1 , (a, b) ∈ H}| < γr/2] ≤ |S2 | · exp(−γ 2 r/2) < 1.
S1

42
Thus, there exists a set S1 of size r such that for every b ∈ S2 we have

|{(a, b) : a ∈ S1 , (a, b) ∈ H}| ≥ γr/2.

A.2 Properties of the desired interpolating polynomial

As preparation of the interpolating step, we start by describing the desired monomial support of
the interpolating polynomials and observing some of their properties.
Definition A.3. For n, N, p ∈ N such that p ≥ 2, the sets Nd,D and Nd,D,p are defined as follows.

• Nd,D is the set of exponent vectors of trivariate monomials with (1, 1, d)-degree bounded by D,
i.e.,

Nd,D := (i, j, k) ∈ Z3≥0 : i + j + dk ≤ D .

• Nd,D,p is the set of exponent vectors of trivariate monomials xi y j z k with (1, 1, d)-weighted
degree at most D such that the z degree is either zero, or is not a multiple of p. More formally,

Nd,D,p := (i, j, k) ∈ Z3≥0 : i + j + dk ≤ D and (k is either zero or p ∤ k) . ♦

The following simple claims now give us bounds on the sizes of the sets Nd,D and Nd,D,p . We
defer the proofs of the claims to Appendix A.4.

Claim A.4. For all d, D ∈ N, we have

(D 3 /3d − 5/2D 2 + Dd/6) ≤ |Nd,D | ≤ (D 3 /3d + 3/2D 2 + Dd/6) .

Claim A.5. For all d, D ∈ N, with p ≥ 2, we have |Nd,D,p | ≥ |Nd,D |/2.

Finally, we combine the bounds in Claim A.4 and Claim A.5 to get the following claim.
D3
Claim A.6. For all d, D, p ∈ N such that D > 20d and p ≥ 2, |Nd,D,p | ≥ 12d .

Interpolation
We now rely on the above estimates to interpolate a low-degree polynomial that explains the
function value in the points table on a constant fraction of points. But first, we recall some
notation: H, S1 , S2 be as guaranteed by Lemma A.2 and let r be the size of S1 . Define S =

(a, b) ∈ F2 : b ∈ S2 , (a, b) ∈ H . From Lemma A.2, we have that |S| ≥ γ · q 2 . With this nota-
tion in place, we have the following lemma.

43
Lemma A.7. Let F be a finite field of characteristic p and let D be a natural number satisfying
|Nd,D,p | > r(D + 1). There exists a non-zero polynomial A(x, y, z) ∈ F[x, y, z] with deg1,1,d (A) ≤ D
such that

• For every a ∈ S1 , we have

A(a, y, Pℓx=a (y)) = 0

as a univariate polynomial in y, where Pℓx=a is the best-fit degree-d polynomial for f the line
ℓx=a .

• If p is the characteristic of the field F, then, A is supported on monomials of the form xi y j z k

where k is either zero or is not divisible by p.
P
Proof. Let A(x, y, z) = i,j,k∈Nd,D Ai,j,k xi y j z k , where Ai,j,k are indeterminates. For each a ∈ S1 ,
the constraint A(a, y, Pℓx=a (y)) = 0 can be viewed as (D + 1) homogeneous linear constraints on the
indeterminates Ai,j,k obtained by equating the coefficients of y 0 , y 1 , . . . , y D to zero. Note that since
the (1, 1, d)-weighted degree of every monomial in A is at most D, the degree of A(a, y, Pℓx=a (y)) is
at most D. Hence, as long as |Nd,D,p | > r·(D+1), there must exist a non-zero solution to the system
of equations, and therefore a non-zero polynomial (1, 1, d)-degree D that is entirely supported on
monomials from the set Nd,D,p and therefore satisfies the second item in the lemma by definition.

We now infer that any polynomial satisfying the conditions in Lemma A.7 imply that it satisfies
many more vanishing conditions.

Lemma A.8. Suppose A(x, y, z) is a polynomial with deg1,1,d (A) ≤ D such that for all a ∈ S1 we
have A(a, y, Pℓx=a (y)) = 0. If D < γ/2 · |S1 |, then we have that

A(a, b, f (a, b)) = 0 for all (a, b) ∈ S = {(a, b) : b ∈ S2 , (a, b) ∈ H} .

Proof. Fix a b ∈ S2 and let Rb (x) = Pℓy=b and Q(x) = A(x, b, Rb (x)). Suppose a ∈ S1 with
(a, b) ∈ H, then Rb (a) = Pℓx=a (b) = f (a, b) and hence

A(a, b, Rb (a)) = A(a, b, Pℓx=a (a)) = 0.

Since deg(Q) ≤ deg1,1,d (A) ≤ D, and |{(a, b) : a ∈ S1 } ∩ H| ≥ γ/2 · |S1 | = γr/2, the condition that
D < γr/2 implies that Q(x) is identically zero. Since Q(x) = A(x, b, Rb (x)) = 0, for any a such that
(a, b) ∈ H, we have that A(a, b, Rb (a)) = A(a, b, f (a, b)) = 0 as claimed.

44
A.3 Proof of the interpolation lemma
We now have the necessary ingredients to prove Theorem 4.3.

Proof of Theorem 4.3. Let H be the set guaranteed by Claim A.1 and γ ∈ (0, 1] such that H =
2γ · q 2 = Ω(ε2 · q 2 ); without loss of generality let the two directions guaranteed by Claim A.1 be the
standard axes.
Let r = 900d/γ 2 and D = γr/3 = 300d/γ . From these choice of parameters, we have the following
inequalities.
• D > 20d
(3003 d2 )
• |Nd,D,p | ≥ D 3 /12d = 12γ 3

6(300)2 d2
• r(D + 1) ≤ 2rD = γ3
450d
• γr/2 = γ

Thus, we have that D ≥ 20d, |Nd,D,p | > r(D + 1) and D < γr/2.
Now, Instantiating Lemma A.2 with this choice of r, we obtain sets S1 , S2 and let S be defined
as

S = (a, b) ∈ F2 : b ∈ S2 , (a, b) ∈ H .

By construction, |S| ≥ γ · q 2 = Ω(ε2 · q 2 ). Since S ⊆ H and each x ∈ S satisfies εx ≥ ε/2 (from

P
Claim A.1), we have that x∈S εx = Ω(ε3 · q 2 ).
Since D 3 /12d > r(D + 1) and D < γr/2, we have by Lemma A.7 and Lemma A.8 that there is
a non-zero polynomial Ã(x, y, z) with deg1,1,d (Ã) ≤ D, entirely supported on monomials in the set
Nd,D,p such that Ã(a, b, f (a, b)) = 0 for all (a, b) ∈ S. In order to proceed further in the proof, we
need the following claim whose proof we defer to the end of this section.

Claim A.9. The polynomial à depends on the variable z, i.e., there exists a monomial of the form
xi y j z k with non-zero coefficient such that k is non-zero.

Since à depends on z and is only supported on monomials where the z degree is either zero
or is not a multiple of p, we have from the first item of Proposition 2.4 and the linearity of Hasse
derivatives that ∂¯z (Ã) must be non-zero. We take A(x, y, z) to be a non-zero polynomial of the
minimum (1, 1, d)-weighted degree such that it vanishes on (a, b, f (a, b)) for all (a, b) ∈ S and ∂¯z (A)
is non-zero. We know that the polynomial à interpolated above is one such polynomial. Therefore,
deg(1,1,d) (A) ≤ deg(1,1,d) (Ã) ≤ D. The polynomial A thus obtained satisfies the hypothesis of
Lemma 2.9, and therefore, by Lemma 2.9, we get that Discz (A) must be non-zero.
This proves all the properties of A and S claimed by Theorem 4.3.

To complete the proof of Theorem 4.3, we now prove Claim A.9.

45
Proof of Claim A.9
Proof of Claim A.9. If à does not depend on z at all, then it is a non-zero bivariate polynomial
of (1, 1)-weighted degree, i.e., total degree at most D, and it vanishes at all points (a, b) in the set
 ofzeroes of à on F × F can be at most
S ⊆ F × F. But, from Lemma 2.2, we have that the number
Dq, which by the choice of D is at most 300/γ · dq ≤ O dq
ε2
. We also have that |S| > Ω(ε2 q 2 ).
Thus, for any sufficiently large constants c1 , c2 , we have that if q > c1 · d/εc2 , then, |S| exceeds
O(dq/ε2 ), thereby implying that à must be identically zero, which is a contradiction. Thus, à must
depend on z.

A.4 Proofs of Claim A.4, Claim A.5 and Claim A.6

Proof of Claim A.4. From the definition, we have that

⌊D/d⌋ (D−kd) (D−kd−i)

X X X
|Nd,D | ≥ 1,
k=0 i=0 j=0
⌊D/d⌋
X
≥ (D − kd)(D − kd + 1)/2 ,
k=0
⌊D/d⌋
X
≥ 1/2 · (D − kd)2 ,
k=0
⌊D/d⌋
X
≥ 1/2 · (D 2 + k2 d2 − 2Ddkd), ,
k=0
 
⌊D/d⌋ ⌊D/d⌋ ⌊D/d⌋
X X X
≥ 1/2 · ( D2 ) + ( k2 d2 ) − ( 2Ddk)
k=0 k=0 k=0

For ease of notation, let D̃ denote ⌊D/d⌋. So, we have (D/d−1) ≤ D̃ ≤ (D/d). Now, from standard
estimates on sums and sums or squares of the first n natural numbers, we have the following.

D̃
X
D 2 ≥ D 2 · (D̃ + 1) ∈ [D 3 /d, D 3 /d + D 2 ] .
k=0

Similarly,
 
D̃
X XD̃
k2 d2 = d2 ·  k2  ≥ d2 · (D̃)(D̃ + 1)(2D̃ + 1)/6 .
k=0 k=0

46
Plugging in the bounds for D̃, we get that

D̃
X
k2 d2 ∈ [1/6(2D 3 /d − 3D 2 + Dd), 1/6(2D 3 /d + 3D 2 + Dd)] .
k=0

Finally,

D̃
X
2Ddk = 2Dd · D̃(D̃ + 1)/2 ,
k=0

which, again using the upper and lower bounds on D̃ satisfies

D̃
X
2Ddk ∈ [D 3 /d − D 2 , D 3 /d + D 2 ] .
k=0

Putting the estimates together, we get the claim.

Proof of Claim A.5. For any d, D ∈ N, and k ∈ Z≥0 , let Ak be the set defined as

Ak = (i, j) ∈ Z2≥0 : i + j + dk ≤ D .

Clearly, the size of Ak is a non-increasing function of k. Thus, for every integer ℓ ≥ 1 and for every
j ∈ {1, 2, . . . , p − 1} we have that |Aℓp | < |A(ℓ−1)p+j |, thereby implying that
 
p
X
1
|Aℓp | < |A(ℓ−1)p+j | .
p
j=1

Summing over ℓ, we get that

 
⌊D/d⌋ ⌊D/d⌋ p
X 1 X X
|Aℓp | < |A(ℓ−1)p+j | .
p
ℓ=1 ℓ=1 j=1

P⌊D/d⌋
We now note from the definitions of the sets Nd,D and Ak that |Nd,D | = k=0 |Ak | and, thus, we
have that
⌊D/d⌋
X 1
|Aℓp | < |Nd,D | .
p
ℓ=1

47
Moreover, the size of the set Nd,D,p satisfies

⌊D/d⌋
X  
1
|Nd,D,p | ≥ |Nd,D | − |Aℓp | ≥ 1− |Nd,D | ,
p
ℓ=1

which, for p ≥ 2 gives |Nd,D,p | ≥ 12 |Nd,D |.

Proof of Claim A.6. From the lower bound on |Nd,D | in Claim A.4, we have that |Nd,D | ≥ D 3 /3d −
5/2D 2 . Now, if D/d > 20, then |Nd,D | ≥ D 2 (D/3d − 5/2) ≥ D 3 /6d. Combining this with the
bound in Claim A.5 completes the proof.

B Variations on Low-degree Testing

In this section we show the equivalence of some standard variations of low-degree testing. These
results are essentially folklore - our proofs follow closely the proofs from Arora and Sudan [AS03,
Section 2.2], confirming along the way that the proofs continue to hold even when the field size is
only linear in the degree.
(f,d)
For a function f : Fm → F and line ℓ in Fm , let Pℓ be the best fit degree d univariate
polynomial for f on the line ℓ. We now, define the following quantities for any function f : Fm → F,
line ℓ and plane π

(f,d)
δf (ℓ) := Pr [Pℓ (x) 6= f (x)],
x∈ℓ
(f,d)
δf (π) := E [δf (ℓ)] = Pr [Pℓ (x) 6= f (x)],
ℓ∈π ℓ∈π
x∈ℓ
(f,d)
δf := E[δf (π)] = Prm [Pℓ (x) 6= f (x)].
π ℓ∈F
x∈ℓ

As mentioned in Remark 4.1, the general low-degree test is provided two oracles f : Fm
q → Fq
and P : L (m) → F≤d m
q [t], the first mapping points in Fq to values in Fq and the latter mapping lines
in Fm
q to (a table of evaluations of) a degree ≤ d univariate polynomial. The probability that the
low-degree test accepts is given by Prx,ℓ∋x [f (x) = P [ℓ](x)].
We say x is β-good for (f, P ) if Prℓ∋x [P [ℓ](x) = f (x)] ≥ β. Note that if the low-degree test
accepts (f, P ) with probability β then at least (β/2) fraction of the points are (β/2)-good for (f, P ).
Conversely, if the low-degree test accepts (f, P ) with probability at most β = β1 β2 then at most β1
fraction of the points are β2 -good for (f, P ) (for any choice of β1 , β2 satisfying β = β1 β2 ).
Definition B.1 (Weak form of LDT). Given field Fq , integer parameters m ≥ 1, d ≥ 0, real β > 0
and function h : R>0 → R>0 , we say that weak low-degree testing holds for (Fq , m, d, β0 , h) if the
following is true:

48
 For all f : Fm
q → Fq and P : L
(m) → F [t]≤d and all β > β , if (f, P ) pass the low-
q 0
degree test with probability at least β, then there exists some degree d polynomial Q such
that agree(f, Q) ≥ h(β).

♦
Note that Theorem 4.2 asserts that there are universal constants C, α such that weak low-degree
testing holds for (Fq , 2, d, β0 , h) for h : β 7→ αβ 4 provided q ≥ Cd/β07 .

There are two incomparable ways to strengthen a weak low-degree test and we first define the
list-decoding variant.
Definition B.2 (List-decoding form of the LDT). Given a field Fq , integer parameters m ≥ 1,
d ≥ 0, real β0 > 0 and function h1 : R>0 → R>0 , we say that list-decoding low-degree testing holds
for (Fq , m, d, β0 , h1 ) if the following is true:

For all β > β0 , f : Fm

q → Fq and P : L
(m) → F [t]≤d , there is a (possibly empty) list of
d
2
at most C ≤ h1 (β) m-variate degree d polynomials Q1 , . . . , QC such that agree(f, Qi ) ≥
h1 (β) for each i, and

Pr [x is β-good for (f, P ) and f (x) ∈

/ {Q1 (x), . . . , QC (x)}] ≤ β0 .
x∈Fm
q

The above must hold regardless of the LDT acceptance probability of (f, P ). Indeed, if (f, P ) passes
the LDT with probability less than β · β0 , then the list of polynomial may be empty as there can be
at most β0 fraction of β-good points for (f, P ). ♦
Another strengthening of the weak LDT is in terms of the agreement probability.
Definition B.3 (High-agreement form of the low-degree test). Given a finite field Fq , integer pa-
rameter m ≥ 1, d ≥ 0, real β0 > 0 and a function h2 : R>0 → R>0 , we say that (Fq , m, d, β0 , h2 ) if
the following is true:

For every β > β0 , f : Fm

q → Fq and P : L
(m) → F [t]≤d such that (f, L) passes the low-
q
degree test with probability at least β, then there exists an m-variate degree d polynomial
Q such that agree(f, Q) ≥ β − h2 (β0 ).

♦
In our applications we will assume h, h1 , h2 are monotone non-decreasing functions. We will see
that a weak low-degree test with h(β) → 0 as β → 0 imply both the list-decoding variant and the
high-agreement variant for some appropriate functions h1 , h2 that also satisfy h1 (β), h2 (β) → 0 as
β → 0.
In our applications we will assume h, h1 , h2 are monotone non-decreasing functions. Further we
assume h(β) → 0 as β → 0 and show that h1 (β) → 0 and h2 (β) → 0 as β → 0. Note that the two

49
implications above are incomparable and do not directly imply each other. However as we will see
in the proof, the implication in (1) is useful to prove (2).

Lemma B.4 (Weak LDT implies list-decoding LDT). Suppose Fq , m, d, β0 and h(·) are such that
weak low-degree testing holds for (Fq , m, d, β0 , h). Then list-decoding low-degree testing holds for
(Fq , m, d, β0′ , h1 ) for
!
√ e
• β0′ satisfying β0′ > β0 and h(β0′2 ) ≥ max √ q ,
4q, 2(d+1)
q
, dq

• h1 : β 7→ 1/2 · h(β · β0 ).

Proof. Fix any β > β0′ . Let Q1 , . . . , QC be the set of all m-variate degree d polynomials that have
agreement at least η = 1/2 · h(β · β0′ ) > 1/2 · h(β0′2 ) with f on Fm
q . By the Johnson bound, we have
p ′
that C ≤ 2/η since η ≥ 2 d/q by the choice of β0 .
Consider the following randomly chosen function g : Fm q → Fq given by g(x) = f (x) if x 6∈ ∪i Si
and g(x) ∼ Unif(Fq ) if x ∈ ∪i Si . Claim B.5 below asserts that with positive probability we have
that agree(g, Q) < 2η for every degree d polynomial Q. Fix a g such that this holds. By the
weak low-degree test applied to g (in contrapositive form) it follows that (g, P ) pass the low-degree
test with probability at most β · β0′ since β · β0′ > β0′2 ≥ β0 and h(β · β0′ ) = 2η. Therefore,
Prx∈Fm
q
[x is β-good for (g, L)] < β0′ . But now note that x ∈
/ ∪i Si satisfies f (x) = g(x) and so such
an x is τ -good for (g, P ) iff it is β-good for (f, P ). We conclude that

Pr [x is β-good for (f, L) and f (x) ∈

/ {P1 (x), . . . , PC (x)} ]
x∈Fm
q

= Prm [x is β-good for (g, L) and f (x) ∈

/ {P1 (x), . . . , PC (x)} ]
x∈Fq

= Prm [x is β-good for (g, L)] ≤ β0′ .

x∈Fq

To complete the proof, we only need to prove the following claim.

Claim B.5. If η ≥ max{ √eq , 4(d+1)

q } then

m
Pr[∃Q : Fm
q → Fq , deg(Q) ≤ d , agree(g, Q) ≥ 2η] ≤ q
−(η/4)q
< 1.

Proof. Fix a degree d polynomial Q. If Q ∈

/ {Q1 , . . . , QC } then agree(Q, f ) < η and
if Q ∈ {Q1 , . . . , QC } we have for every x ∈
/ ∪i Si , Q(x) 6= f (x) = g(x). Thus in either
/ ∪i Si |Q(x) = g(x) = f (x)}| ≤ ηq m . Thus to have agreement at least
case we have |{x ∈
2η with Q, g must satisfy |{x ∈ ∪i Si |P (x) = g(x)}| ≥ ηq m . We show below that the
m
probability that this happens is at most q −(η/2)q .
For a fixed set S ⊆ ∪i Si of size ηq m , the probability that g and Q agree on the set is
m qm

ηq m ≤ q (η/2)q m (using η ≥ e/√q).
q −ηq . The number of sets S is at most ηq m ≤ (e/η)

50
 We conclude that the probability that there exists a set S ⊆ ∪i Si of size at least ηq m
m
such that g and Q agree on S is at most q −(η/2)q .
Now to conclude the proof we take a union bound over all Q’s. The number m vari-
ate monomials of degree at most d is clearly at most (d + 1)m and so the number of
m m m
polynomials is at most q (d+1) ≤ q ((η/4)q) ≤ q (η/4)q (where the first inequality uses
η ≥ 4(d + 1)/q and the second uses m ≥ 1). We conclude that the probability that
there exists Q of degree at most d such that |{x ∈ ∪i Si |Q(x) = g(x)}| > ηq m is at most
m
q −(η/4)q . The claim follows.

This completes the proof of Lemma B.4.

Lemma B.6 (Weak LDT implies high-agreement LDT). Suppose Fq , m, d, β0 and h(·) are such that
weak low-degree testing holds for (Fq , m, d, β0 , h). Then, high-agreement low-degree testing holds for
(Fq , m, d, β0′′ , h2 ) for

• β0′′ satisfying β0′′ > β0′ and β0′′3 · h1 (β0′′2 ) > 2/q and β0′′ · h1 (β0′′2 ) ≥ 2d/q,
• h2 : β 7→ 3β,

where β0′ and h1 (·) are as implied by Lemma B.4.

In other words, if (f, P ) passes the LDT with probability β > β0′′ , then there is some m-variate
degree d polynomial Q such that agree(f, Q) ≥ β − 3β0′′ .

Proof. By applying Lemma B.4 (with β = β0′′ ≥ β0′ ) we have that for any pair (f, P ), there exist at
most C = 2/h1 (β0′′ ) polynomials Q1 , . . . , QC , each with agreement at least h1 (β0′′ ) with f such that
 
Prm x is β0′′ -good for (f, P ) and f (x) ∈
/ {Q1 (x), . . . , QC (x)} ≤ β0′ .
x∈Fq

Let Si = {x : Qi (x) = f (x)} and let ηi = q −m |Si |. Assume w.l.o.g. that η1 ≥ · · · ≥ ηC . Note that
we are not guaranteed that C ≥ 1 and the above list of polynomials be empty. However, we wish
to show that if (f, P ) passes the LDT with probability β ≥ β0′′ , then C ≥ 1 and η1 ≥ β − γ, where
γ = 3β0′′ . We will do so by proving that the acceptance probability of the low-degree test on (f, P )
is upper bounded by η1 + γ.
Define a line ℓ to be standard if P [ℓ] ∈ {Q1 |ℓ , . . . , QC |ℓ }. Say the ℓ is abnormal if there exists
i ∈ [C] such that |{x ∈ ℓ : Qi (x) = f (x)}| ≥ (η1 + β0′′ ) · q. For a non-standard line ℓ say that a
point x ∈ ℓ is coincidental for ℓ if there exists i ∈ [C] s.t. Qi (x) = P [ℓ](x). Finally say that a pair
(x, ℓ) with x ∈ ℓ is unexplained if f (x) = P [ℓ](x) and f (x) 6∈ {Q1 (x), . . . , QC (x)}.
We now upper bound the probability that the low-degree test accepts a random pair (x, ℓ) by
consider various cases. We first note that for the low-degree test to accept a pair (x, ℓ) at least one
of the following must happen:

51
 1. ℓ is standard and normal and the low-degree test accepts, or
2. ℓ is abnormal, or
3. ℓis non-standard and x is coincidental for ℓ, or
4. (x, ℓ) is unexplained.

The typical case is Item 1 where ℓ is standard and normal. In this case the probability over x that
the low-degree test accepts the pair (x, ℓ) is at most η1 + β0′′ .
For Item 2, the probability that a random line ℓ is abnormal is upper bounded by γ1 := C/(γ 2 q)
by a Chebychev argument (for fixed i ∈ C the expected fraction of agreement is ηi and a random line
contains q pairwise independent random samples of points from Fm ′′ ′′3 ′′2
q ). Since β0 satisfies β0 ·h1 (β0 ) >
2/q, we this probability is bounded by β0′′ .
For Item 3, the probability that a point x on a non-standard line ℓ is coincidental for the line
is at most Cd/q (for every i ∈ [C] there are at most d points where Qi (x) = P [ℓ](x)). Since β0′′
satisfies β0′′ · h1 (β0′′2 ) > 2d/q, we this probability is bounded by β0′′ as well.
And finally for Item 4, the probability that a pair (x, ℓ) is unexplained is, by the list-decoding
version of the LDT, at most β0′ ≤ β0′′ .
We thus conclude that (f, P ) passes the low-degree test accepts with probability at most η1 +
3β0′′ .

52